CN106406516A - Local real-time movement trajectory characteristic extraction and identification method for smartphone - Google Patents

Abstract

A method for feature extraction and recognition of local real-time motion trajectory of a smart phone, which is divided into two stages of training and recognition. In the training stage, the user carries the smart phone to do different actions and collects the data of the three-axis acceleration sensor; The feature points of the valley are then encoded in binary; after encoding, the feature points are vectorized; and the actions with the same peak and valley values are matrixed, and multiple samples are collected for training on the host computer to establish a standard library of user action features. In the identification stage: transplant the user action standard library to the mobile phone, the user carries the mobile phone to make actions, collect sensor data on the mobile phone end, and the smartphone performs feature point extraction locally, matches the user action standard library, and then recognizes the user action. The feature vectors required to be stored in the invention for identifying the motion trajectory of the mobile phone are binary and have a small data scale, the identification process is relatively simple and does not require a large number of calculations, and is suitable for embedded devices such as smart phones with limited resources.

Description

A local real-time motion trajectory feature extraction and recognition method for smart phones

technical field

The invention relates to the technical field of smart phone track recognition, and specifically refers to a local real-time motion track feature extraction and recognition method for smart phones.

Background technique

The built-in acceleration sensor of the mobile phone can be used to detect the magnitude and direction of the acceleration of the mobile phone according to the acceleration sensor, and then perceive the state in the three-dimensional space. For the mobile phone acceleration sensor, there are coordinates in the three directions of X, Y, and Z. When the user carries the mobile phone to do different actions, different values will be generated for the three coordinate axes of X, Y, and Z. The data generated by the three axes at the same time reflects the user's status. At present, the motion trajectory feature extraction and recognition process of smartphones is relatively complicated, and the computing power of smartphones is weak, and the storage space is small, and additional equipment is needed for non-real-time motion trajectory feature extraction and recognition, which cannot meet the local real-time requirements of smartphone devices. for the identification of motion trajectories.

Invention content:

The present invention proposes a local real-time feature extraction and recognition method for the motion trajectory of a smart phone. The feature point vector stored in the method to identify the motion trajectory of the mobile phone is binary and the data scale is small. The recognition process is relatively simple and does not require a large number of calculations. Embedded devices such as smartphones are restricted.

For this reason, the adopted technical scheme is:

A method for feature extraction and recognition of local real-time motion trajectory of a smart phone, the method is divided into two stages of training and recognition; in the training stage: the user carries the smart phone to do different actions, and collects and separates the three-axis acceleration sensor data of the smart phone; Extract the feature points of peaks and valleys from user gestures and behaviors; perform binary code quantization on the feature points of peaks and valleys; vectorize the feature points after encoding; matrixize the actions with the same number of peaks and valleys, collect multiple samples for training on the host computer, and establish user Action feature standard library; in the recognition stage: transplant the user action feature standard library to the smart phone, when the user carries the smart phone to do different actions, collect sensor data on the mobile phone end, open up the buffer area on the mobile phone end, extract motion trajectory data, and establish multiple The thread extracts segmented feature points, matches the standard library of user action features, and the smartphone performs feature point extraction locally in real time, and then recognizes user actions.

Compared with the complex process of feature extraction and recognition of the motion trajectory of smart phones, and the weak computing power and small storage space of smart phones, it is necessary to use additional equipment to perform non-real-time feature extraction and recognition of motion trajectory. The identification has the following advantages: (1) It can identify the motion trajectory for smart phones or some small embedded devices. (2) The calculation of feature points is simple, and the normalization processing of feature points is reduced, which is suitable for smart phones with weak computing power. (3) After the feature points of the motion trajectory are quantized and encoded, the feature point vectors required to identify the motion trajectory of the mobile phone are binary and the data size is small, which is suitable for embedded devices such as smartphones with limited storage resources.

Description of drawings:

Fig. 1 is the flow chart of smart phone motion track training of the present invention;

Fig. 2 is the flow chart of smart phone motion track recognition of the present invention;

Fig. 3 is the data waveform diagram that smart phone sensor of the present invention produces;

Fig. 4 is the coding diagram of separating sensor data and feature extraction in the present invention;

Fig. 5 is that the present invention utilizes the window to detect peak and valley figure;

Fig. 6 is a peak and valley feature vectorization diagram of the present invention;

Fig. 7 is a matrix diagram of peak and valley characteristics of special actions in the present invention.

detailed description:

The present invention will be described in further detail below in conjunction with the accompanying drawings.

The schematic diagram of the training process of the present invention is as shown in Figure 1:

First, the user carries a smart phone or a small embedded device to do different actions. The three-axis acceleration sensor generates data, separates the data according to the X, Y, and Z axes, and then searches for the peak and valley feature points in each direction and encodes them. The feature points are vectorized, and for gestures with the same number of peaks and troughs, they are matrixed, multiple samples are collected, and the artificial neural network is used to train the samples to form a standard library of user action features.

Separate the smartphone acceleration sensor along the X, Y, and Z axes. When a certain action has its uniqueness in the X-axis direction, it can be recognized. If it cannot be identified on the X axis, there is no need to compare the Y and Z axes. And process the Y axis and Z axis in turn. Using the method can reduce data processing at the mobile phone end and save computing resources of the smart phone.

Use the form of a piece of data segmentation to extract peak and valley feature points for user gesture behavior. This method provides a method for extracting feature points of peaks and valleys of data. The method is simple, the amount of calculation is small, and the key feature points can be found by filtering noise points.

The coded peak and valley feature points are expressed in the form of vectors, and the feature points can be formatted through vectorization, which is convenient for further calculation and can improve calculation efficiency.

Actions with the same number of peaks and valleys are processed in a matrix. Movements with the same number of crests and troughs form a The matrix, M represents the number of peaks and valleys on the three coordinate axes. Computational efficiency is further improved.

The schematic diagram of the identification process of the present invention is shown in Figure 2:

Firstly, the user action feature standard library is transplanted into the storage space of the smart phone. When the user carries a smart phone to do different actions, collect sensor data on the mobile phone side, develop a buffer zone on the mobile phone side, extract motion trajectory data, establish multi-threaded extraction of segmented feature points, match user action feature standard library, and local real-time on the smart phone Feature point extraction is performed to identify user actions.

The process will be further described below.

As shown in Figure 3: when the user makes an action, the acceleration sensor of the smartphone generates data in the three directions of X, Y, and Z, processes the data, and then stitches them together to form three waveforms. The data on the coordinate axis of the mobile phone reflects the user's motion state.

As shown in Figure 4: According to the X, Y, and Z directions of the smartphone sensor, the three sets of sensor data of the acceleration sensor are processed separately, and the order of the peaks and valleys is used as the characteristic point of the waveform. The detected feature points are encoded in such a way that the peak is 1 and the valley is 0. Utilizing binary encoding saves computer storage resources.

The method of peak and valley feature extraction:

As shown in Figure 5: use the sliding window to identify the peaks and troughs of a piece of data. A fixed number of generated data is used as a fixed-value sliding window, and the user's motion state data is all in the fixed-value sliding window, and the sliding window is divided into equal parts of small sliding windows, and the detection of peak and valley data is carried out in this small sliding window. There will be peaks and troughs in each small sliding window, sorted according to the order of appearance, compare the peaks and troughs, and take the first few peaks with the largest value and the last few troughs with the smallest value as the value of each axis Feature points. Through this method, some noise points can be filtered out and key feature points can be found.

As shown in Figure 6: the feature points obtained from the peaks and valleys are vectorized. The peaks and valleys of the data generated by the three sets of acceleration sensors are generated in time order to form three sets of vectors, and the data in each vector represents the feature points in each direction and the order in which the feature points appear.

As shown in Figure 7: the data formed by the special action is matrixed. The data generated by the separation sensor in all directions can be processed in a matrix if the peak and valley lengths of the three coordinate axes are the same. forming one matrix.

Recognition stage: Smartphone sensors will generate data every moment, and there will be synchronization problems between the calculation of data peaks and valleys and the generation of data, and the use of open buffers and multi-threading to solve synchronization problems.

The implementation method is as follows: In the process of peak and valley detection, a large sliding window buffer is defined, and the data generated by the sensor is continuously iteratively stored in a In the fixed-length matrix, where N is based on a length value set by the length of the display window, the latest data is added to the matrix, while the data stored earlier is discarded, and the length of the matrix is maintained at a fixed value N, and then the three fixed-length matrices Equally divided, open up multiple threads for the detection of peak and valley features on the small sliding window buffer after equalization.

In the process of comparing the user's motion state with the sample library, the X-axis data is first compared, and if a certain action has its uniqueness in the X-axis direction, it can be identified. If it cannot be identified on the X axis, there is no need to compare the Y and Z axes. Process the Y axis and Z axis in turn.

The effect of the present invention is further illustrated through specific application scenarios below:

Scenario 1: Human-computer interaction field. For the current human-computer interaction, it is necessary to purchase supporting equipment, which is relatively expensive. As a highly integrated device, a smart phone has many sensors inside, which generate a large amount of data all the time. Moreover, the smart phone has weak computing power and small storage space, and can carry out key points for the generated data. storage. Recognize user actions, and then map other functions for human-computer interaction.

Scenario 2: Normalization processing of feature points. For some waveform feature processing, sometimes there will be a lot of interference, such as speed, force, and the collected feature points need to be normalized. Through this method, it is not necessary to perform normalization with the same large or small scale. , only need to find the peak and valley feature points according to the time sequence.

Scenario 3: Quantization and calculation of feature points. For some specific fields, after the feature points are found, the feature points cannot be calculated. By matrixing the feature points in this paper, the feature points can be further processed.

To sum up, the present invention effectively solves the relatively complex problem of extracting motion trajectory features of smart phones through a series of processing such as data separation, feature extraction, feature encoding, and feature point vectorization on the smart phone acceleration sensor, and the waveform The features are encoded and quantified, and the key information is effectively stored, which greatly saves the storage resources of the smart phone. The feature encoding can reduce the normalization processing of the feature points, save the computing resources of the smart phone, and do not need to use additional equipment. , which can perform real-time motion trajectory feature extraction and recognition locally, and is suitable for resource-constrained smart phones and other embedded devices.

Claims (4)

1. A local real-time motion trajectory feature extraction and recognition method for smart phones, characterized in that: the method is divided into two phases of training and recognition; in the training phase: the user carries a smart phone to do different actions, and collects and separates smart phone three Axial acceleration sensor data; extract peak and valley feature points for user gesture behavior; perform binary code quantization on peak and valley feature points; vectorize feature points after encoding; matrixize actions with the same number of peaks and valleys, and collect multiple samples in the upper position machine for training, and establish a user action feature standard library; in the recognition stage: transplant the user action feature standard library to the smartphone. Motion trajectory data, establish multi-threaded extraction of segmented feature points, match the standard library of user action features, and perform feature point extraction locally on the smartphone in real time, and then recognize user actions. 2. The local real-time motion trajectory feature extraction and recognition method of a kind of smart phone according to claim 1, characterized in that: the smart phone acceleration sensor is separated along three axes. 3. The local real-time motion trajectory feature extraction and recognition method of a smart phone according to claim 1, characterized in that: the peak and valley feature points are extracted to the user's gesture behavior in the form of a section of data segmentation. 4. The local real-time motion track feature extraction and identification method of a smart phone according to claim 3, wherein: the form of using a section of data segmentation is specifically: using a sliding window to identify the peak and trough of a section of data , use a fixed number of data generated as a fixed-value sliding window, and the user's motion state data is all within the fixed-value sliding window, and the sliding window is divided into equal parts of the sliding window on average, and the peak and valley data are processed in this small sliding window Detection, there will be peaks and troughs in each small sliding window, sort them in the order of appearance, compare the peaks and troughs, take the first few peaks with the largest value and the last few troughs with the smallest value, as each Axis feature points.