CN108399362A - A kind of rapid pedestrian detection method and device - Google Patents

Abstract

The invention discloses a kind of rapid pedestrian detection method and devices, and described method includes following steps：Step S1 builds the configurable depth model based on convolutional neural networks, learns the network parameter for structure using training sample, obtains the model for test process；Step S2, input test sample, the changing rule for being perceived domain using neural network by trained model is detected the target object within the scope of different scale using different middle layers, predict the block diagram of target object in image, the present invention perceives the changing rule in domain by using neural network, the target object within the scope of particular dimensions is detected using different middle layers, the relationship in perception domain and article size has preferably been adapted to, has effectively increased testing result.

Description

A fast pedestrian detection method and device

technical field

The invention relates to the technical field of pedestrian detection, in particular to a deep learning-based fast pedestrian detection method and device for embedded systems.

Background technique

As a part of object detection in computer vision, pedestrian detection is of great significance in real-world applications. With the maturity of image acquisition technology and the decline of storage technology costs, more and more cameras are deployed in public places. On the other hand, With the implementation of autonomous driving and intelligent transportation, car cameras have also produced massive video resources. Traditional manual screening and processing is not only inefficient and consumes a lot of manpower and material resources, but also may introduce some human factors, resulting in some deviations. In recent years, deep learning has made unprecedented breakthroughs in the field of computer vision. Not only is the efficiency far superior to that of manpower, but its accuracy also exceeds that of humans in many fields. Therefore, the subject of effective use of deep learning methods for pedestrian detection has attracted much attention.

People are one of the most important targets in video surveillance or automatic driving, and the primary task of pedestrian detection is to identify the existence of human bodies and provide corresponding annotation information. Due to the uneven quality of images captured in the real world, the detection of small objects and occluded objects has always been a difficult point in pedestrian detection. On the other hand, vehicle cameras often capture some blurred images. In such images There are also a large number of pedestrian-like objects that are not pedestrians. As for embedded systems, since large-scale neural network models with strong recognition capabilities are usually difficult to efficiently run on embedded devices with limited computing resources, and the application requirements for embedded devices are real-time, so the detection accuracy And efficiency is the top priority for fast pedestrian detection for embedded systems.

Contents of the invention

In order to overcome the deficiencies in the above-mentioned prior art, one object of the present invention is to provide a fast pedestrian detection method and device, by using the change rule of the neural network perceptual domain, using different intermediate layers to detect target objects within a specific scale range Detection, better adapted to the relationship between the perception domain and the size of the object, effectively improving the detection results.

Another object of the present invention is to provide a kind of fast pedestrian detection method and device, obtain the squeeze VGG-16 network that adapts to the requirement of embedded system by adjusting and training the network of VGG-16, effectively reduce the parameter quantity of network model and accelerate Computational efficiency.

Another object of the present invention is to provide a fast pedestrian detection method and device, which can amplify the feature map of a specific network layer through deconvolution, which enhances the detection of small objects. Compared with the traditional image enlargement method, Almost no increase in video memory and calculation.

Yet another object of the present invention is to provide a fast pedestrian detection method and device, by using an area 1.5 times the size of the target object as the background semantic feature and adding it to the network, it has excellent detection of fuzzy objects and long-distance small objects performance.

In order to achieve the above and other purposes, the present invention proposes a rapid pedestrian detection method, which includes the following steps:

Step S1, constructing a configurable deep model based on convolutional neural network, using training samples to learn the constructed network parameters, and obtaining a model for the testing process;

Step S2, input test samples, use different intermediate layers to detect target objects in different scale ranges through the trained model, and use the changing rules of the neural network perceptual domain to predict the block diagram of the target object in the image.

Preferably, step S1 further includes:

Build configurable convolutional neural network-based deep models;

input training samples;

Initialize the convolutional neural network and its parameters, including the weights and biases of each layer connection in the network layer;

The forward propagation algorithm and the backward propagation algorithm are used to learn the constructed network parameters by using the training samples, that is, the model used for the testing process.

Preferably, the depth model includes a multi-scale target candidate network and a target detection network, and the target candidate network is based on the differences in features proposed by different layers of the convolutional neural network, and generates candidates for target objects of different scales in the middle layer. Block diagram: the target detection network performs refined classification and detection on the basis of the candidate block diagram output by the target candidate network.

Preferably, the convolutional neural network is formed by stacking convolutional layers, downsampling layers, and upsampling layers. The convolution layer refers to performing a convolution operation on the input image or feature map in a two-dimensional space to extract hierarchical features; the downsampling layer uses a max-pooling operation without overlap, which is used to extract shape and Offset features are constant, while reducing the size of the feature map and improving computational efficiency; the upsampling layer refers to the deconvolution operation on the input feature map in two-dimensional space to increase the pixels of the feature map .

Preferably, the depth model adopts the Squeeze VGG-16 convolutional neural network as the backbone network, and the Squeeze VGG-16 convolutional neural network adopts the conv1-1 layer and the following 12-layer Fire module layer as feature extraction network structure.

Preferably, the target candidate network is based on the Squeeze VGG-16 convolutional neural network, according to the convolutional layer features, in Fire9, Fire12, conv6 and the added pooling layer, to generate network branches to detect different scales Regression of object proposals.

Preferably, on the basis of the target candidate area, the target detection network uses the image area of the preset multiple size of the target candidate area as the background semantic information of the target, and performs an upsampling of the feature map of the Fire9 layer as an enhanced pair The information perceived by small objects, and the background semantic information and upsampling information are pooled in the region of interest to obtain fixed-size features, and then a fully connected layer is added to perform category and final candidate frame regression.

Preferably, the training samples include RGB image data and label information of the pedestrian area in the image, and the image data for actual training is a small patch cut out according to the area where the pedestrian is located.

Preferably, the backward propagation algorithm needs to first obtain the loss function between the target block diagram predicted by the forward propagation and the actual target block diagram of the image Then obtain its gradient to the parameter W, and use the gradient descent algorithm to update W to minimize the loss function Assume that there are M branches in the middle layer that can output the target candidate area, l ^m represents the loss function of branch m, α _m represents the weight of the l ^m function, S={S ¹ , S ² ,...,S ^M } refers to the target of the corresponding scale object, the loss function can be defined as:

To achieve the above purpose, the present invention also provides a rapid pedestrian detection system, including:

The training unit is used to build a configurable deep model based on convolutional neural network, use training samples to learn the constructed network parameters, and obtain a model for the testing process;

The detection unit is used to input test samples, and use different intermediate layers to detect target objects in different scale ranges by using the trained model and using the changing law of the neural network perceptual domain, and predict the block diagram of the target object in the image.

Compared with the prior art, a fast pedestrian detection method and device of the present invention learn from the method of compressing the network, adjust and train the VGG-16 network to obtain a squeeze VGG-16 network that meets the requirements of the embedded system, and effectively reduce the network model. parameter quantity and speed up the calculation efficiency; on the other hand, in view of the problem that the perception domain and the size of the object are inconsistent in the traditional detection method, the present invention utilizes the change rule of the neural network perception domain (that is, the deeper the neural network layer, the larger the perception domain, suitable for Detect larger target objects), use different intermediate layers to detect target objects within a specific scale range, better adapt to the relationship between the perception domain and the size of the object, and effectively improve the detection results; in addition, in order to enhance the detection of small objects detection, the present invention uses a deconvolution method to amplify the feature map of a specific network layer, compared to the traditional image magnification method, which hardly increases the amount of video memory and calculation; in order to enhance the detection of fuzzy objects, in this layer On the feature map, the area 1.5 times the size of the target object is used as the background semantic feature to add to the network. It has excellent performance for the detection of fuzzy objects and long-distance small objects.

Description of drawings

Fig. 1 is a flow chart of the steps of a fast pedestrian detection method of the present invention;

Fig. 2 is the schematic diagram of Squeeze VGG-16 neural network structure in the specific embodiment of the present invention;

Fig. 3 is the schematic diagram of Fire module in the specific embodiment of the present invention;

4 is a schematic structural diagram of a target candidate network in a specific embodiment of the present invention;

5 is a schematic structural diagram of a target detection network in a specific embodiment of the present invention;

6 is a schematic diagram of the process of rapid pedestrian detection in a specific embodiment of the present invention;

Fig. 7 is a system architecture diagram of a fast pedestrian detection device of the present invention;

Fig. 8 is a detailed structural diagram of a training unit in a specific embodiment of the present invention;

Fig. 9 is a detailed structure diagram of a detection unit in a specific embodiment of the present invention.

Detailed ways

The implementation of the present invention is described below through specific examples and in conjunction with the accompanying drawings, and those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific examples, and various modifications and changes can be made to the details in this specification based on different viewpoints and applications without departing from the spirit of the present invention.

FIG. 1 is a flow chart of steps of a rapid pedestrian detection method of the present invention. As shown in Figure 1, a kind of fast pedestrian detection method of the present invention comprises the following steps:

Step S1, constructing a configurable deep model based on convolutional neural network, using training samples to learn the constructed network parameters, and obtaining a model for the testing process. In a specific embodiment of the present invention, the depth model is composed of two sub-networks: the first sub-network is a multi-scale target candidate network, which is used to extract character features and give candidate areas, specifically, the target candidate network Based on the differences in features proposed by different layers of the convolutional neural network, candidate block diagrams for pedestrians of different scales are generated in the middle layer; the second sub-network is the target detection network, which enhances the detection effect, and it shares parameters with the target candidate network. Refined classification and detection are performed on the basis of candidate block diagrams. Specifically, step S1 further includes:

Step S100, constructing a configurable deep model based on a convolutional neural network.

The convolutional neural network is formed by stacking a convolutional layer, a downsampling layer, and an upsampling layer. The convolutional layer refers to performing a convolution operation on an input image or feature map in a two-dimensional space to extract hierarchical features. ; The down-sampling layer uses a max-pooling operation without overlapping, which is used to extract features with invariant shape and offset, while reducing the size of the feature map and improving computational efficiency; the up-sampling layer refers to the The input feature map is deconvoluted in two-dimensional space to increase the pixels of the feature map, which is mainly used in the target detection network to improve the detection effect. In the specific embodiment of the present invention, Squeeze VGG-16 volume The convolutional neural network is used as the backbone network. As shown in Figure 2, the Squeeze VGG-16 convolutional neural network uses the conv1-1 layer and the following 12-layer Fire module as the convolutional layer to extract features; the pool1 -pool5 is the downsampling layer; use the pre-trained model on the ImageNet dataset as initialization. That is, the present invention first uses the ImageNet data set to pre-train Squeeze VGG-16 as network initialization.

Fig. 3 is a schematic structural diagram of a Fire module in a specific embodiment of the present invention. As shown in Figure 3, the Fire module consists of two convolution layers with a convolution kernel size of 1×1 and a convolution layer with a convolution kernel size of 3×3. The purpose is to use a convolution kernel of 1×1 instead of 3×3 convolution kernel, which reduces the amount of parameters by 9 times, but in order not to affect the representation ability of the network, it is not all replaced, but part of it uses a 1×1 convolution kernel, and part of it uses a 3×3 convolution Kernel, another advantage of this is to reduce the input channels of the 3×3 convolution kernel, and at the same time reduce the amount of parameters. Specifically, the Fire module first uses a 1×1 convolution layer to perform dimensionality reduction operations on the input layer , and then referring to the GoogLeNet structure, use 1×1 and 3×3 convolutional layers to extract features, and finally connect the two parts of the features, which greatly reduces the amount of calculation and model parameters.

Fig. 4 is a schematic diagram of an architecture of a target candidate network in a specific embodiment of the present invention. In a specific embodiment of the present invention, the target candidate network is based on the Squeeze VGG-16 convolutional neural network, and according to the convolutional layer characteristics, a total of 4 layers are generated in Fire9, Fire12, conv6 and the added pooling layer to generate network branches, The branch performs the regression of the candidate boxes of detected objects at different scales. But for the Fire-9 layer, it is relatively close to the lower layer of the backbone network. Compared with other layers, it will have a greater impact on the gradient, and the learning process is unstable, so there is an additional buffer (buffer) layer, as shown in the det-conv layer in Figure 4. As shown, the buffer layer prevents the gradient of the detection branch from being directly back-propagated (backpropagated) to the backbone layer.

The present invention utilizes the change law of the neural network perception domain (that is, the deeper the neural network layer, the larger the perception domain, which is suitable for detecting larger target objects), and uses different intermediate layers to detect target objects within a specific scale range, which is better It adapts to the relationship between the perception domain and the size of the object, and effectively improves the detection results.

Fig. 5 is a schematic diagram of the architecture of the target detection network in a specific embodiment of the present invention. The target detection network shares parameters with the target candidate network, and summarizes the candidate frames of the target candidate network to enhance the ability of the monitoring network to distinguish objects from backgrounds. In a specific embodiment of the present invention, the target detection network, on the basis of the target candidate area, uses a picture area 1.5 times the size of the target candidate area as the background semantic information of the target; performs an upsampling of the feature map of the Fire9 layer, As the information to enhance the perception of small objects, the background semantic information and upsampling information are pooled in the region of interest (ROI pooling) to obtain fixed-size features, and then a fully connected layer is added to perform category and final candidate frame regression. , specifically, the backbone cnn layer is connected to a node of proposals, which is used to summarize the candidate frame information obtained by the target candidate network; on the other hand, for the feature map of the fire9 layer, W and H are the width and height of the input image, cube 1 Represents the mapping of the object area on the feature map, and cube 2 represents the mapping of the context area on the feature map. The context area is about 1.5 times the object area. At the same time, in order to strengthen the detection of small objects, an upsampling is performed on the Fire9 layer , and then similar to the faster RCNN algorithm, use the pooling of the region of interest to obtain fixed-size features; connect (concat) the features processed by the Fire9 layer with the features summarized by the proposals, and then add a fully connected layer to classify and the regression of the final candidate box, which will not be described here.

Step S101, input training samples.

The training process needs to provide the corresponding frame of the reference person in the image. At the same time, in order to speed up the training, the training process cuts out the image containing the reference person from the original image to form a patch (image block), and the patch is smaller than the original image. , used for training, effectively speeding up the training process. Specifically, in the present invention, the input training sample includes RGB image data and label information of the pedestrian area in the image, and the image data used for actual training is a small patch (image block) obtained by cutting out the area where the pedestrian is located. Expressed in mathematical language, training samples Where X _i represents a patch of the training picture; in practical applications, in addition to the category of pedestrians, there are other categories, such as background, cyclists, sitting people and other K categories, so the labeled data Y _i = (y _i , b _i ) consists of class labels y _i ∈ {0, 1, 2, ..., K} and box plot coordinate points composed of is the starting coordinate point of the upper left corner of the block diagram, is the boxplot width and height.

Step S102, initialize the convolutional neural network and its parameters, including the weight and bias of each layer connection in the network layer. Specifically, the present invention uses the ImageNet dataset to pre-train the Squeeze VGG-16 convolutional neural network as network initialization.

Step S103, using the forward propagation algorithm and the backward propagation algorithm, using the training samples to learn the constructed network parameters, that is, the model used in the testing process.

In the present invention, the forward propagation algorithm first normalizes the size of the input image to 3×480×640, intercepts a patch with a size of 3×448×448 and corresponding label information as the input of the convolutional neural network, After convolutional layer, downsampling layer and rectified linear unit layer (ReLU Nonlinearity Layer), in Fire9 layer, the image feature map size is 512×60×80; in Fire12 layer, the feature map size is 512×30×40, in the back The size of the feature maps of the two branches are 512×15×20 and 512×8×10, respectively. On different feature maps, the four coordinate points and category information of the target block diagram are obtained by convolution. Taking the Fire9 layer as an example, assuming that only pedestrians and backgrounds are detected, the output is a feature size of 6×60×80, of which 6 Contains two categories of background, pedestrian and four coordinate points of the candidate frame. In the target detection network, the candidate block diagrams obtained by each branch layer are summarized at the proposals node, and at the same time, they are superimposed with the background semantic information of the Fire9 layer and the features obtained by the pooling operation of the region of interest through the upsampling information to make the final block diagram Regression and categorical regression.

In the present invention, the backward propagation algorithm needs to first obtain the loss function of the target block diagram predicted by the forward (ie forward) propagation and the actual target block diagram of the image Then obtain its gradient to the parameter W, and use the gradient descent algorithm to update W to minimize the loss function Assuming that the middle layer has M branches that can output target candidate areas (the perceptual domain of M scales can approximately detect all target objects in the image), l ^m represents the loss function of branch m, α _m represents the weight of l ^m function, S = {S ¹ , S ² ,..., S ^M } refers to the target object of the corresponding scale, then the loss function can be defined as:

The loss function, for a specific detection layer m, only contributes to the loss function if the target scale is within the range that m can detect, so the loss function is defined as

Among them, p(X)=(p ₀ (X), ..., p _K (X)) represents the probability distribution of the target category; λ is the balance coefficient; b is the 4 coordinate points of the frame diagram, Refers to the coordinate points obtained by forward propagation; in the loss function, the cross-entropy loss function is used to define the category regression, ie

L _cls (p(X), y)=-log _y (P(X)) (3)

Use the smooth Manhattan distance criterion (smooth L1 criterion) to perform the regression of the target block diagram, defined as follows

In step S2, the trained model is used to detect target objects in different scales using different intermediate layers according to the change law of the neural network perceptual domain, and predict the block diagram of the target object (such as a pedestrian) in the image.

Specifically, step S2 further includes:

Step S200, loading the trained model;

Step S201, inputting test samples;

Step S202 , using the trained model, using different intermediate layers to detect pedestrians in different scales according to the changing rules of the neural network perceptual domain, and predict the block diagram of the pedestrian in the image. 6 is a schematic diagram of the process of fast pedestrian detection in a specific embodiment of the present invention, that is, using the target candidate network in the model on the basis of the Squeeze VGG-16 convolutional neural network, according to the convolutional layer features, in fire9, fire12, conv6 and increase The pooling layer of the pooling layer has a total of 4 layers to generate network branches, and detects target candidate areas of objects at different scales (middle layer a, middle layer b, middle layer c); The image area 1.5 times the size of the candidate area is used as the background semantic information of the target, and the feature map of the Fire9 layer is up-sampled once as information to enhance the perception of small objects, and the background semantic information and up-sampled information are pooled in the area of interest Obtain fixed-size features, and then add a fully connected layer to perform category and final candidate box regression. Preferably, in step S202, a deconvolution method is also used to enlarge the feature map of a specific network layer.

The pedestrian detection method proposed by the present invention uses two evaluation indexes for reference respectively: the average precision rate mAP and the number of frames per second FPS. mAP is used to evaluate the intersection ratio between the final detection area and the real target person area, and the average precision rate under different intersection ratios; FPS is mainly an efficiency indicator, which refers to the number of pictures that can be processed per second.

FIG. 7 is a system architecture diagram of a rapid pedestrian detection device according to the present invention. As shown in Figure 7, a fast pedestrian detection device of the present invention includes:

The training unit 70 is configured to construct a configurable deep model based on a convolutional neural network, use training samples to learn constructed network parameters, and obtain a model for the testing process. In a specific embodiment of the present invention, the depth model constructed by the training unit 70 is composed of two sub-networks: the first sub-network is a multi-scale target candidate network, which is used to extract character features and give candidate regions, specifically , the target candidate network is based on the difference of features proposed by different layers of the convolutional neural network, and generates candidate block diagrams for pedestrians of different scales in the middle layer; the second sub-network is the target detection network, which enhances the detection effect. The network shares parameters, and performs refined classification and detection on the basis of candidate block diagrams. Specifically, as shown in Figure 8, the training unit 70 further includes:

A model construction unit 701, configured to construct a configurable deep model based on a convolutional neural network.

The convolutional neural network is formed by stacking a convolutional layer, a downsampling layer, and an upsampling layer. The convolutional layer refers to performing a convolution operation on an input image or feature map in a two-dimensional space to extract hierarchical features. ; The down-sampling layer uses a max-pooling operation without overlapping, which is used to extract features with invariant shape and offset, while reducing the size of the feature map and improving computational efficiency. The up-sampling layer refers to the The input feature map is deconvoluted in two-dimensional space to increase the pixels of the feature map. In a specific embodiment of the present invention, the Squeeze VGG-16 convolutional neural network is used as the backbone network.

In a specific embodiment of the present invention, the target candidate network is based on the Squeeze VGG-16 convolutional neural network, and according to the convolutional layer features, a total of 4 layers are generated in fire9, fire12, conv6 and the added pooling layer to generate network branches, The branch performs the regression of the candidate boxes of detected objects at different scales. But for the fire-9 layer, it is closer to the lower layer of the backbone network. Compared with other layers, it will have a greater impact on the gradient, and the learning process is unstable. Therefore, a buffer (buffer) layer is added. The buffer layer prevents the gradient of the detection branch from being Directly back-propagated (back propagation) to the backbone layer.

The target detection network shares parameters with the target candidate network, and summarizes the candidate frames of the target candidate network to enhance the ability of the monitoring network to distinguish objects from backgrounds. In a specific embodiment of the present invention, the target detection network, on the basis of the target candidate area, uses a picture area 1.5 times the size of the target candidate area as the background semantic information of the target; performs an upsampling of the feature map of the Fire9 layer, As the information to enhance the perception of small objects, the background semantic information and upsampling information are pooled in the region of interest to obtain fixed-size features, and then a layer of fully connected layers is added to perform the regression of categories and final candidate boxes. Specifically, The backbone cnn layer is connected to a proposal subnet, W and H are the width and height of the input image, cube 1 represents the pooling of the object area, and cube 2 represents the pooling of the context area, the context area is about 1.5 times the object area, and for Strengthen the detection of small objects, and then perform an upsampling on the Fire9 layer, and then similar to the faster RCNN algorithm, use the pooling of the region of interest to obtain fixed-size features, and then add a layer of fully connected layers to perform category and final candidate boxes return.

The training sample input unit 702 is configured to input training samples.

Specifically, the training samples Where X _i represents a patch of the training picture, and the labeled data Y _i = (y _i , b _i ) consists of the category label y _i and the frame coordinate points composition.

The initialization unit 703 is used to initialize the convolutional neural network and its parameters, including the weight and bias of each layer connection in the network layer. Specifically, the present invention uses the ImageNet dataset to pre-train the Squeeze VGG-16 convolutional neural network as network initialization.

The sample training unit 704 is configured to use the forward propagation algorithm and the backward propagation algorithm to learn the constructed network parameters, that is, the model used in the testing process, by using the training samples.

In the present invention, the forward propagation algorithm first normalizes the size of the input image to 3×480×640, intercepts a patch with a size of 3×448×448 and corresponding label information as the input of the convolutional neural network, After convolutional layer, downsampling layer and rectified linear unit layer (ReLU Nonlinearity Layer), in Fire9 layer, the image feature map size is 512×60×80; in Fire12 layer, the feature map size is 512×30×40, in the back The size of the feature maps of the two branches are 512×15×20 and 512×8×10, respectively. On different feature maps, the four coordinate points and category information of the target block diagram are obtained by convolution. Taking the Fire9 layer as an example, assuming that only pedestrians and backgrounds are detected, the output is a feature size of 6×60×80, of which 6 Contains two categories of background, pedestrian and four coordinate points of the candidate frame. In the target detection network, the candidate block diagrams obtained by each branch layer are summarized at the proposals node, and at the same time, they are superimposed with the background semantic information of the Fire9 layer and the features obtained by the pooling operation of the region of interest through the upsampling information to make the final block diagram Regression and categorical regression.

The backward propagation algorithm needs to first obtain the loss function of the target block diagram predicted by the forward propagation and the actual target block diagram of the image Then obtain its gradient to the parameter W, and use the gradient descent algorithm to update W to minimize the loss function Assuming that the middle layer has M branches that can output target candidate areas (the perceptual domain of M scales can approximately detect all target objects in the image), l ^m represents the loss function of branch m, α _m represents the weight of l ^m function, S = {S ¹ , S ² ,..., S ^M } refers to the target object of the corresponding scale, then the loss function can be defined as:

The loss function, for a specific detection layer m, only contributes to the loss function if the target scale is within the range that m can detect, so the loss function is defined as

Wherein, p(X)=(p ₀ (X), . . . , p _K (X)) is the probability distribution of the target category. In the loss function, the category regression is defined using the cross-entropy loss function, namely

L _cls (p(X), y)=-log _y (P(X))

Use the smooth L1 criterion to perform the regression of the target block diagram, which is defined as follows

The detection unit 71 is used to input test samples, and uses different intermediate layers to detect target objects (such as pedestrians) in different scale ranges by using the trained model and using the change law of the neural network perceptual domain, and predicts the target object in the image ( such as a block diagram of a pedestrian).

Specifically, as shown in FIG. 9, the detection unit 71 further includes:

A model loading unit 710, configured to load a trained model;

A test sample input unit 711, configured to input a test sample;

The image prediction unit 712 is configured to use the trained model to detect pedestrians in different scale ranges by using the trained model and using different intermediate layers to predict the block diagram of pedestrians in the image. Specifically, the image prediction unit 712 uses the target candidate network in the model, based on the Squeeze VGG-16 convolutional neural network, and according to the convolutional layer characteristics, a total of 4 layers in Fire9, Fire12, conv6 and the added pooling layer generate network branches , to detect the target candidate areas of the object at different scales; then use the target detection network, on the basis of the target candidate area, use the image area 1.5 times the size of the target candidate area as the background semantic information of the target, and perform the feature map of the Fire9 layer One upsampling, as the information to enhance the perception of small objects, the background semantic information and upsampling information are pooled in the region of interest to obtain fixed-size features, and then add a fully connected layer to perform category and final candidate frame regression .

In summary, a method and device for fast pedestrian detection in the present invention draws on the method of compressing the network, adjusts and trains the network of VGG-16 to obtain a squeeze VGG-16 network that meets the requirements of the embedded system, and effectively reduces the amount of parameters of the network model And speed up the calculation efficiency; on the other hand, in view of the problem that the perception domain and the size of the object are inconsistent in the traditional detection method, the present invention utilizes the change rule of the neural network perception domain (that is, the deeper the neural network layer, the larger the perception domain, which is suitable for detecting large objects). Some target objects), use different intermediate layers to detect target objects within a specific scale range, better adapt to the relationship between the perception domain and the size of the object, and effectively improve the detection results; in addition, in order to enhance the detection of small objects , the present invention uses the deconvolution method to enlarge the feature map of a specific network layer. Compared with the traditional image enlargement method, it hardly increases the amount of video memory and calculation; in order to enhance the detection of fuzzy objects, the feature map of this layer In the above, the area 1.5 times the size of the target object is used as the background semantic feature to add to the network, and it has excellent performance for the detection of fuzzy objects and long-distance small objects.

The above-mentioned embodiments only illustrate the principles and effects of the present invention, but are not intended to limit the present invention. Any person skilled in the art can modify and change the above-mentioned embodiments without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention should be listed in the claims.

Claims (10)

1. A fast pedestrian detection method, comprising the steps of: Step S1, constructing a configurable deep model based on convolutional neural network, using training samples to learn the constructed network parameters, and obtaining a model for the testing process; Step S2, input test samples, use different intermediate layers to detect target objects in different scale ranges through the trained model, and use the changing rules of the neural network perceptual domain to predict the block diagram of the target object in the image. 2. A kind of fast pedestrian detection method as claimed in claim 1, is characterized in that, step S1 further comprises: Build configurable convolutional neural network-based deep models; input training samples; Initialize the convolutional neural network and its parameters, including the weights and biases of each layer connection in the network layer; The forward propagation algorithm and the backward propagation algorithm are used to learn the constructed network parameters by using the training samples, that is, the model used for the testing process. 3. A kind of fast pedestrian detection method as claimed in claim 2, it is characterized in that, described this depth model comprises multi-scale target candidate network and target detection network, and described target candidate network proposes based on different layers of convolutional neural network According to the difference of features, candidate frame diagrams of target objects of different scales are generated in the middle layer; the target detection network performs refined classification and detection on the basis of the candidate frame diagrams output by the target candidate network. 4. a kind of fast pedestrian detection method as claimed in claim 3 is characterized in that: described convolutional neural network is formed by stacking of convolution layer, down-sampling layer, up-sampling layer, and described convolution layer refers to The input image or feature map is convolved in two-dimensional space to extract hierarchical features; the downsampling layer uses a max-pooling operation without overlap, which is used to extract features with invariant shape and offset, and at the same time Reduce the size of the feature map and improve the calculation efficiency; the upsampling layer refers to the deconvolution operation on the input feature map in two-dimensional space to increase the pixels of the feature map. 5. A kind of fast pedestrian detection method as claimed in claim 4, is characterized in that: described depth model adopts Squeeze VGG-16 convolutional neural network as backbone network, and described Squeeze VGG-16 convolutional neural network adopts conv1- The first layer and the following 12-layer Fire module layer are the network structure of feature extraction. 6. A kind of fast pedestrian detection method as claimed in claim 5, is characterized in that: described target candidate network is on the basis of described Squeeze VGG-16 convolutional neural network, according to convolution layer feature, in Fire9, Fire12, Conv6 and the added pooling layer generate network branches for the regression of candidate frames of objects detected at different scales. 7. A kind of fast pedestrian detection method as claimed in claim 5, it is characterized in that: described target detection network is based on described target candidate area, uses the picture area of preset multiple size of target candidate area as the background of target Semantic information, the feature map of the Fire9 layer is upsampled once as information to enhance the perception of small objects, and the background semantic information and upsampling information are pooled in the region of interest to obtain a fixed-size feature, and then add a layer of full The connection layer performs the regression of the category and the final candidate box. 8. A kind of fast pedestrian detection method as claimed in claim 1, is characterized in that: described training sample comprises RGB image data and the mark information of pedestrian area in image, and the image data that actual training is used is to obtain according to the area cropping of pedestrian's place A small patch. 9. A kind of fast pedestrian detection method as claimed in claim 1, it is characterized in that: described backpropagation algorithm, needs to first obtain the loss function of the target frame diagram of forward propagation prediction and the actual target frame diagram of image Then obtain its gradient to the parameter W, and use the gradient descent algorithm to update W to minimize the loss function Assume that there are M branches in the middle layer that can output the target candidate area, l ^m represents the loss function of branch m, α _m represents the weight of the l ^m function, S={S ¹ , S ² ,...,S ^M } refers to the target of the corresponding scale object, the loss function can be defined as: 10. A rapid pedestrian detection system comprising: The training unit is used to build a configurable deep model based on convolutional neural network, use training samples to learn the constructed network parameters, and obtain a model for the testing process; The detection unit is used to input test samples, and use different intermediate layers to detect target objects in different scale ranges by using the trained model and using the changing law of the neural network perceptual domain, and predict the block diagram of the target object in the image.