WO2016129091A1 - Object detection system and object detection method - Google Patents

Abstract

An object detection method which is characterized by comprising: a camera that is disposed on a mobile unit; a GPS that acquires image position information, which is information on the position where an image was captured by the camera; a first storage unit that stores the image and the image position information; a second storage unit that stores, as information on an object to be detected, an object image characteristic quantity, which is the image characteristic quantity of the object to be detected, and installation position information; an image selection unit that selects the image as a selection image on the basis of the image position information and the installation position information and determines a specified region in the selection image as a detection region; a detection unit that performs object detection by determining the presence or absence of the object to be detected in the detection region using the degree of similarity between the image characteristic quantity extracted from the detection region and the object image characteristic quantity; and an output unit that gives a notification to a user if it is determined that there is no object to be detected.

Description

Object detection system and object detection method

The present invention relates to an object detection apparatus and an object detection method for detecting an object from an image.

With the expansion and aging of roads and railroad track networks, a larger area has to be regularly maintained, and the necessary cost is increasing. In the case of railways, not only maintenance of tracks and sleepers, but also maintenance of signs, railroad crossings, traffic lights, etc. is necessary for safe operation management. Signs can break or become difficult to see due to disasters such as typhoons or changes over time.

This is the same on the road where the vehicle runs. For example, there is a problem that the road surface display disappears as time passes. For this reason, it is necessary to perform maintenance work such as confirming that there is no problem through regular audits, and re-installing if any problem is found. However, it is expensive even if the maintenance worker periodically confirms. Based on the management ledger created at the time of installation, it is very time consuming to check the correctness of the installation visually while traveling.

On the other hand, in Patent Document 1, “a video image taken by the video camera 1 while the vehicle is running is recorded on the video tape 2 and reproduced on the video deck 4 to be automatically recognized by the road sign recognition device 10 ( It is disclosed that a road sign is detected in real time by a personal computer).

JP 2009-217832 A

However, Patent Document 1 assumes that only a sign having a fixed shape is detected and updated. For this reason, a process of estimating an approximate shape from the existence probability of the contour line and recognizing it as a sign is performed. However, assets such as signs actually installed outdoors are often discolored or deformed due to various conditions such as the weather, and there is a problem that detection accuracy is lowered.

In order to solve the above problems, for example, the configuration described in the claims is adopted. The present application includes a plurality of means for solving the above-described problem. To give an example, an object detection system is an image that is information about a camera provided on a moving body and a position where the camera has captured an image. The target image feature quantity, which is the image feature quantity of the target object, and the installation position information are stored as information about the target object to be detected. A second storage unit that performs selection, an image selection unit that selects an image as a selection image based on the image position information and the installation position information, and determines a specified region in the selection image as a detection region; A detection unit that performs object detection by determining whether or not an object exists using the similarity between the image feature amount extracted from the detection region and the target image feature amount, and it is determined that the object does not exist If And having and an output unit for performing notification to THE.

Alternatively, in the object detection method, a first step of storing, in the first storage unit, an image captured by a camera provided on a moving body and image position information that is information on a position where the camera has captured the image; A second step of storing, in the second storage unit, target image feature quantities, which are image feature quantities of the target objects, and installation position information indicating positions where the target objects are installed, as information relating to the target object to be detected; Based on the information and the installation position information, a third step of selecting an image as a selection image and determining a designated area in the selection image as a detection area, and whether or not an object exists for the detection area The fourth step of performing object detection by making a determination using the similarity with the target image feature amount extracted from the detection area, and notification to the user when it is determined that the target does not exist I do And having a step, the.

According to the present invention, it is possible to improve the accuracy of detecting the presence or absence of assets (objects).

It is a figure which shows the whole structure of this invention. It is a figure which shows the outline | summary of the road asset management system of this invention. It is a figure which shows the physical hardware constitutions of this invention. It is a figure which shows the whole processing flow of this invention. It is a figure which shows an example of the asset information of this invention. It is a figure which shows an example of the process area | region of this invention. It is a figure which shows the detection process of this invention. It is a figure which shows an example of the position estimation process of this invention. It is a figure which shows the determination process of this invention. It is a figure which shows an example of the screen display of this invention. It is a figure which shows an example of the asset management system of this invention. It is a figure which shows an example of the captured image of this invention on a time series. It is a figure which shows the redetermination process of this invention. It is a figure which shows an example of the detection part of this invention. It is a figure which shows an example of the detection target object of this invention.

FIG. 1 is a diagram showing the configuration of the present invention. The camera 100 is installed on a moving body and acquires images before and after the moving direction of the moving track (or road). The GPS receiver 120 is a receiver that acquires GPS information, and can thereby acquire position information of a moving moving body.

The video storage unit (first storage unit) is connected to the camera 100 and the GPS receiver 120, acquires video from the camera, GPS information from the GPS receiver 120, and records this. At this time, the camera 110 and the GPS receiver 120 transmit time stamps to the acquired video and GPS information, respectively, and then transmit them to the video storage unit 130. In the video storage unit, the position information is added and recorded for each frame image constituting the video by synchronizing with both time stamps.

The asset information DB 144 stores asset images and installed location information. The asset feature DB 145 stores search image feature amounts extracted from asset images. In the present specification, these may be collectively referred to as a second storage unit.

The video confirmation unit 140 detects an asset in the video stored in the video storage unit 130, confirms the status of the asset, and notifies the asset. Below, each part which comprises the image | video confirmation part 140 is demonstrated. In the process selection unit 141, the asset information DB 141 is collated based on the position information of the video stored in the video storage unit 130, and the frame image to be detected is selected, and the content of the process to be performed in the detection unit 142 is selected. To do. The detection unit 142 performs image recognition processing for detecting assets from the video stored in the video storage unit according to the selection made by the processing selection unit 141.

When the position estimation unit 143 detects a certain asset in the detection unit, the position estimation unit 143 estimates how far the detected asset is from the camera. By adding the estimated relative position information of the asset and the GPS information, the actual position where the detected asset is installed is estimated. The detection result is output to the determination unit 147 and stored in the detection result DB 146. In the detection result DB 146, information on the detected asset is recorded.

The determination unit 147 compares the asset position obtained by the detection unit 142 and the position estimation unit 143 with the asset information DB 144 to determine whether the detected asset is installed at the correct position. If an abnormality such as no asset is determined, the result is output to the notification unit 148.

The notification unit 148 notifies the user by using voice, video, or the like when the determination unit 147 finds an abnormality. When displaying a video, the video is acquired from the video storage unit 130 or the detection result DB (third storage unit) 146 and displayed.

FIG. 2 is an outline of a moving body 200 in which the object detection device of the present invention is installed. In the present embodiment, the moving body is described as an example of a train, but it goes without saying that the present invention can also be applied to a vehicle running on a road. FIG. 2A shows an example in which all components are installed in a moving body. The image confirmation unit confirms the status of the asset, and notifies the person boarding the moving body of the result. Alternatively, only the abnormal state and the video at that time are notified to the management center through a network such as the Internet communication network. In this case, since only the video when the abnormality occurs and the determination result need to be sent to the management center, the data transmission capacity in real time can be suppressed.

FIG. 2B shows an example in which the video confirmation unit is installed on a server in the management center. In this case, the moving body has only a camera, a GPS, and a video storage unit, and transmits video to the video confirmation unit in the management center using real-time communication via the Internet or data transfer means such as a USB cable. As a result, asset management is performed by the video confirmation unit in the management center in real time or offline.

FIG. 3 is a diagram showing a physical device configuration of the video confirmation unit. The video confirmation is realized as software that operates on one or a plurality of PCs or servers. I / F such as keyboard and mouse, network I / F that performs image acquisition and result notification through the network, CPU that performs detection processing and position estimation processing, storage unit that stores data such as asset information DB and asset feature DB, video A memory that holds a program for each process constituting the confirmation unit and manages temporary data is connected via a bus. When all the on-road asset management devices are mounted in the mobile body as shown in FIG. 2A, it is possible to provide a video storage device in the storage unit and realize it on one PC.

FIG. 4 is a flowchart showing the processing flow of the entire video confirmation unit. First, in S401, a certain sequence of videos is acquired from the video storage unit. In S402, using the position information in the acquired video sequence as a key, the asset information DB is inquired whether there is an asset to be detected in the section of the actual position where the sequence is captured. In S403, if there is no detection target in the video sequence as a result of the inquiry, the process is not performed, and the process returns to S401 to acquire the next video sequence. If it is an area where assets exist in the video sequence, the process proceeds to the next step.

In S404-407, detection processing is performed for all assets that should exist in the video. First, in S404, the processing of asset n is selected. If the detection target is a sign, an area on the right side of the screen is set as a detection area such as an area on the right side of the screen and an area around the line if the line is branched, and an area in the frame image to be detected is determined. In S405, detection processing suitable for the selected asset n is performed. In S407, the process returns to S404 to perform the same process for the next asset.

In S408, it is determined whether or not the asset to be detected is detected as a result of the processing in S403-07, and if it is detected, the position of the asset is estimated in S409. If there is no asset, the process proceeds to S410.

In S409, the asset position is estimated. The installation position in the world coordinate system can be calculated by estimating the installation position of the asset from the image and the speed information of the moving body, and combining the GPS information and the information on the moving direction of the moving body.

In S410, the position of the asset registered in the asset information DB or whether there is an asset in a specific area around it is determined. The specific section here can be determined by multiplying the error value of GPS or camera time measurement by the current moving speed. Although the detection processing time can be shortened by setting the interval as short as possible, it is necessary to set the detection processing not to be performed due to these errors. As another method, the length of the specific section may be determined from an error in the installation position that may occur during installation. This is a value set based on the experience of maintenance so far, so that even if there is a slight error in the installation position of the asset, it can be detected. As a result, if there is an asset, it is determined that there is no problem, and the process returns to S401. However, if there is no asset, there is a high possibility that some kind of abnormality has occurred, and therefore the process proceeds to S411 to notify the user that there is no asset.

Based on the above, the object detection method described in the present embodiment performs first accumulation of an image captured by the camera 110 provided in the moving body 200 and image position information that is information on a position where the camera captured the image. First information accumulated in the unit, information on the target object to be detected, target image feature quantity that is an image feature quantity of the target object, and installation position information indicating a position where the target object is installed in the second storage unit A second step of storing, a third step of selecting an image as a selected image based on the image position information and the installation position information, and determining a designated area in the selected image as a detection area (S403-407); The fourth step of performing object detection by determining whether or not an object exists in the detection area by using the similarity with the target image feature quantity extracted from the detection area (S408). , When the object is determined not to exist, and having a step (S410-411) for performing notification to a user, the.

By taking the above flow, the detection accuracy can be improved by using the position information to identify and detect assets in the section, instead of using all assets in the asset information DB. . Furthermore, since the frame image for performing asset detection and the detection area in the frame image can be limited, the processing time can be reduced. By preparing the image feature amount extracted from various variations of each asset in the asset feature DB 145, even if the same sign A has a change such as a tilt due to the installation problem or a difference in the description contents Can be detected robustly.

Below, the detail of each part is demonstrated.
FIG. 5 is an example of a table 500 held by the asset information DB. Information about what is installed as assets and where is stored. Also, the date on which the installation was confirmed by the worker and the worker was confirmed is stored.

FIG. 6 is an example of a frame image indicating the processing region and the detection target determined by the processing selection unit (image selection unit) 141. In the process selection unit, based on the asset name and the position information accumulated in the asset information DB, the video frame for performing the asset detection process is limited to only those acquired at the peripheral position where the asset to be detected is located. Further, the detection target area is limited such that the area at the right end of the track is a sign and the area on the road is a branch.

FIG. 7 is a diagram showing details of processing of the detection unit. FIG. 7A shows a search window in the detection process, and the search window 709 is moved by the sliding window method to the detection area 708 designated by the process selection unit. The size and aspect ratio of this search window are also determined by the process selection unit according to the asset.

The flow of the detection process is shown in FIG. First, the search window extraction unit 701 cuts out partial images using each search window 709. Next, the feature amount extraction unit 702 extracts image feature amounts from the extracted partial images. Here, an image feature amount extraction process corresponding to the asset to be detected designated by the process selection unit 141 is performed. As the image feature amount, an intensity distribution in the luminance gradient direction, a histogram in RGB or CIELab color system, and the like are extracted. At this time, the same feature quantity extraction method is used for assets whose detection target areas are close in the image, such as the same feature quantity F1 for the signs A and B, and the same feature quantity F2 for the branches and sleepers on the track. Thus, when the signs A and B are installed side by side, it is not necessary to extract different image feature amounts from the same region a plurality of times, and the detection process can be made efficient.

The image search unit 703 determines the presence / absence of an asset by comparing this with the feature amount of the specific asset in the asset feature DB 145. In the asset feature DB, image feature amounts extracted from images of a plurality of signs A are recorded, and an image search process is performed on this, and a distance value with the most similar image is similar to the sign A. Get as a degree. The smaller the distance value is, the higher the similarity is, that is, it is similar.

The search result integration unit 704 integrates the results of the similarity between all the areas obtained in the plurality of search windows 709 and the marker Ano0 image feature amount, and if there is a search window in which the similarity is a certain threshold value or less. , It is determined that there is a sign A in the search window. On the other hand, if there is no image similar to the asset image feature DB 145 of the sign A, it is determined that there is no sign A, and the fact is notified to the user via the notification unit 148.

Based on the above, the object detection system described in the present embodiment includes a camera 110 provided in the moving body 200, a GPS 120 that acquires image position information that is information on a position where the camera 110 has captured an image, A first accumulator 130 that accumulates image position information, and a second accumulator (144, 155) that accumulates an object image feature quantity that is an image feature quantity of the object and installation position information as information about the object to be detected. ), Image position information and installation position information, an image is selected as a selected image, an image selection unit 141 that determines a specified area in the selected image as a detection area, and an object exists for the detection area A detection unit that performs object detection by using the similarity between the image feature amount extracted from the detection region and the target image feature amount, and determines that the target does not exist. If, and it is having and an output unit for performing a notification to the user.

Through the above processing, the detection accuracy can be improved by using the position information to identify and detect the assets in the section instead of using all the assets in the asset information DB. Furthermore, since the frame image for performing asset detection and the detection area in the frame image can be limited, the processing time can be reduced. In addition, by preparing image feature quantities extracted from various variations of each asset in the asset feature DB 145, there are changes such as inclinations due to installation problems and differences in description contents even with the same sign A. Even in this case, detection can be performed robustly.

FIG. 8 is an image diagram showing an example of position estimation processing in the position estimation unit. FIG. 8A shows an example of the result of asset detection by the detection unit. Signs and track branches are detected. Now, since the position of the moving body (the position of the camera that captured the image) is known from the GPS information, in order to know the position of this asset, the relative distance of the asset to the camera may be estimated.

FIG. 8B is a diagram showing an image when the installation position of the asset detected by the image processing is estimated. By using the camera focal length (f) and the camera parameter of the depression angle and the position of the vanishing point detected from within the image, the distance Z from the camera to the asset in the world coordinate system from the vertical position y on the image coordinate system is obtained. It can be calculated. FIG. 8C is an example in which a broken line is drawn on the image for each distance Z having a constant interval. Accordingly, the distance from the camera can be estimated by detecting the lower end position of the detected sign. Furthermore, by performing position estimation using a laser range sensor, a stereo camera, or the like, errors can be reduced more than distance estimation from an image.

FIG. 8D is an example illustrating a sensor signal that can be acquired from a range sensor when a laser range sensor is used as a position estimation device separately from the camera. In a laser range sensor, a laser is blown toward the front and side of a moving body, and the distance is estimated by reflection thereof. For this reason, a sensor signal can be collected as a distance value with respect to the θ direction from the moving body as shown in FIG. From the result of FIG. 8A, it can be seen that the detected sign is in the θ direction with respect to the moving body, and by obtaining the distance value in that direction, the relative distance value from the camera to the sign can be acquired.

FIG. 8E shows an example of a distance image obtained from a stereo camera when a stereo camera is used as the camera. In a stereo camera, two cameras are installed side by side, and the distance value to each pixel in the image can be estimated by obtaining the shift of the image obtained from each camera. In FIG. 8E, the distance is indicated by a gray value, and the distance value can be acquired for each pixel. For this reason, distance can be estimated by acquiring the pixel value in the area of the detected sign.

The relative position of the asset with respect to the moving object (camera) can be estimated by the above method. The absolute position of the asset in the world coordinate system can be estimated from the relative distance thus obtained and the information on the position of the moving body and the direction of the moving body based on the GPS information.

FIG. 9 is a flowchart showing the determination process in the determination unit. In step S901, the assets detected from the detection unit and the position estimation unit and their installation positions are input. Next, in step S902, the asset information DB is collated. If the asset specified in the asset information DB is not detected, the process goes to S903, and the determination result of “no asset” is output. If there is an asset, the process proceeds to S904.

In S904, the position of the detected asset is collated with the asset information DB. In S905, as a result of the collation in S904, it is determined whether or not the positional deviation of the detected asset is equal to or greater than a certain threshold value. If it is below, the process proceeds to S907.

In S907, the distance value is calculated by comparing the detected asset with the past detection result image stored in the detection result DB. In S908, it is determined whether or not the obtained distance value is equal to or greater than a certain threshold value. If the distance value is equal to or greater than the threshold value, a change due to breakage or aging is conceivable. On the other hand, if the distance value is equal to or smaller than the threshold value, it is considered that there is no problem, and “no notification” is set in S910.

According to the above flow, it becomes possible to classify the status of assets in detail. By determining not only the presence / absence of the asset but also the deviation between the information stored in the asset information DB and the actual position, it is possible to notify the user of the possibility that the asset information DB has an error.

FIG. 10 is a diagram showing an example in which the notification unit notifies the user not only by voice but also by screen display. The example of the screen at the time of processing the label | marker A as an asset of a detection target is shown. The determination result notification unit 1001 displays the determination result of the determination unit. The current video display unit 1002 acquires and displays a video of a peripheral area where the asset A is not detected from the video storage unit. The playback control unit 1002 controls playback and stop of the video at this time. Thereby, when an abnormality has occurred in the asset to be detected, it can be confirmed on the video whether the notification result is really correct. In addition, when an abnormality actually occurs, the user can check the status of the asset without going to the shooting site.

In the past video display unit 1004, if there is a previous asset detection result in the detection result DB 146, an image at that time is displayed. As a result, the user can easily compare whether it was wrong in the past audit and what kind of change has occurred compared to the past time point.

The map information display unit 1005 displays the current playback position on the map based on the GPS information of the captured video. As a result, the user can easily grasp at which position the abnormality has occurred in the asset.

FIG. 11 shows an example in which two cameras are used for asset management. Since other configurations are the same as those of the first embodiment, the description thereof will be omitted as appropriate. In the case of the present embodiment, one camera is installed as the peripheral camera 111 so that the periphery of the front or rear of the railway can be widely viewed. The other camera is installed as a road surface camera 112 facing downward from the camera 111 so that only the road surface portion is photographed.

FIG. 11B shows an image taken by the peripheral camera 111. By shooting so that all the signs, tracks, tunnels, etc. next to the track are within the angle of view, the angle of view is adjusted so that all assets such as signs, tunnels, tracks, etc. fit within the screen.

FIG. 11C is an example of an image photographed by the road surface camera 112. FIG. 11C shows an example in which the camera 112 is rotated 90 degrees from an angle of 111 to capture a vertically long image. You can take a picture of the sleepers on the track separated by shooting downward. In addition, on roads, road markings such as speed restrictions and stop lines can be photographed for easy image recognition. As described above, since only the road surface portion is photographed with high resolution, the asset position can be estimated with higher accuracy.

Hereinafter, portions of the processing of the processing selection unit 141 when two cameras are installed will be described, which are different from the first embodiment. The process selection unit 141 determines which video of the two cameras is used based on the GPS information. In a place where there is an asset on the road surface such as a branch of a track, an image captured by the road surface camera 112 is selected as an image used for detection.

Also, it is possible to perform processing for switching the frame rate by using the fact that the peripheral camera 111 can shoot farther away. First, the line branching portion detection process is performed on the video from the peripheral camera 111. When the track branching portion is detected, parameters such as the frame rate of the road surface camera 112 are controlled. Specifically, how many meters an image is taken is set as a parameter, and the frame rate is controlled in accordance with the current speed of the moving body, which is a fluctuation value. For example, in the case of a camera capable of shooting a range of 5 m, it is assumed that an image is to be acquired every 2 m while looking at the margin. At this time, when the moving speed is 4 m / s, the camera is stored at 0.5 second intervals (2 fps), and when the moving speed is 20 m / s, the camera is stored at 0.1 second intervals (10 fps). Control the frame rate. Also, the faster the moving speed, the more blurred the image becomes and the more difficult it is to detect. Therefore, if the preset 2 m acquisition interval is narrowed according to the speed, asset detection accuracy can be further improved.

Since the road surface image is taken only in a narrow section, when the moving speed of the moving body is high, there is a possibility that the image may be blurred or the property may not be taken. Therefore, it is possible to control the camera parameters in accordance with the situation by grasping the target in advance from the road camera image or estimating the current traveling speed. Also, it is possible to prevent omissions by increasing the frame rate only when assets on the road surface come into the road camera.

FIG. 13 is an example of a processing flow of the asset management apparatus that performs redetermination processing according to the third embodiment of the present invention. The main processing flow is the same as that in FIG. 4 of the first embodiment, and thus the description thereof is omitted. However, since S420 to S422 are newly provided, this will be described here.

In S420, if an asset is not found in a specific section where there should be an asset to be detected in S410, it is determined whether this is the first determination process. If it is the first determination process, the process proceeds to S421 to perform the determination process again on the video in the vicinity of the specific section (referred to as the peripheral section), and if the determination is completed again, the process proceeds to S422.

Specified specific section and surrounding section here will be described in detail. For example, as the specific section, a measurement error of about 1-5 m caused by GPS or camera time information is set. On the other hand, around the specific section, an error of about 10 m that occurs when assets are installed is set. In other words, the value in the specific section is a value for correctly detecting in anticipation of the error range of the present apparatus when the asset information DB is correct. On the other hand, the value of the peripheral section is a value for finding an error in this apparatus when there is an error in the asset information DB. In the above case, the specific section is about 1-5 m centering on the position registered in the asset information DB as the installation position. A section excluding a specific section among sections of about 10 m centering on the installation position is a peripheral section.

In S421, an image shot in the section before and after the section with assets determined first is acquired, and the asset detection process is performed again.

As described above, after the detection process for the specific position, the detection process is also performed for the peripheral position. As a result, most assets whose position information is correctly recorded in the asset information DB can be detected in a short processing time. On the other hand, the asset information DB is installed in a location away from the DB due to an error in creating the asset information DB. It is possible to detect existing assets in the second determination process.

FIG. 14 is a diagram showing a configuration of a detection unit that automatically switches DBs and detects assets according to the fourth embodiment of the present invention. A feature is that a search DB selection unit 706 is newly provided. The search DB selection unit 706 selects an asset image to be compared with the image selected for detection from either the asset feature DB 145 or the detection result DB 146 (or both).

In the search DB selection unit, when the sign A photographed at the exact same place is not accumulated in the detection result DB, the search DB selection unit collates with the data of the sign A in the asset image feature DB. On the other hand, if the same point has been audited many times in the past, as a result, if there are multiple images of the sign at the same location, the image of the sign A taken on another day at the same point An image in the search result DB is selected for collation.

FIG. 15 shows an example of fluctuations that can occur in the same type of label. For example, in the case of a sign indicating a station name, the characters in the plate differ depending on the location, and the size of the plate may change accordingly. Further, the shape of the pole to which the plate is fixed may change, or the pole may be tilted depending on the condition of the ground contact surface as shown in FIG.
In order to detect the fluctuation in the same kind of label with high accuracy, the search DB selection unit selects the DB to be collated. In other words, the detection rate can be improved at locations where the vehicle has traveled in the past because it is the same as the same sign.

Also, when the detection result is stored in the detection result DB, the shooting date and time, weather conditions, and the like are stored together. At the time of detection, it is possible to detect with high accuracy even in a place where auditing has not been performed in the past by targeting only images of signs of the same type that were taken in different places but under similar conditions. .

The weather condition is acquired by acquiring the weather information of the shooting location from the GPS information if the video storage unit 130 is connected to the network 210 via the Internet or the like. Accumulate in the accumulator. For example, when the weather information of an image captured at a certain time is “sunny”, by performing detection processing using only images with a “sunny” tag among images stored in the detection result DB, The accuracy of detection processing can be improved.

Also, when the camera 110 has a function such as auto gain control, the gain value used at the time of control can be used as a kind of weather condition information. With respect to this as well, the same effect can be obtained by using only an image in which the difference between the gain value of the photographed image and the gain value of the image accumulated in the detection result DB is equal to or less than the threshold value.

110 ... Camera 111 ... Peripheral camera 112 ... Road surface camera 120 ... GPS receiver,
130: Video storage unit,
140 ... Video confirmation part,
141 ... Processing selection unit 142 ... Detection unit 143 ... Position estimation unit,
144 ... Asset information DB
145 ... Asset feature DB,
146 ... Detection result DB,
147: Determination unit 148 ... Notification unit 200 ... Mobile body 210 ... Network 220 ... Management center 300 ... Video confirmation device 310 ... I / F
320 ... CPU
330 ... Accumulation unit 340 ... Network I / F
350 ... Memory 500 ... Asset information table 701 ... Search window extraction unit 702 ... Feature quantity extraction unit 703 ... Image search unit 704 ... Search result integration unit 705 ... Search result output unit 706 ... Search DB selection unit 1001 ... Determination result notification unit
1002 ... Current video display unit 1003 ... Playback control unit 1004 ... Past video display unit 1005 ... Map display unit

Claims (10)

A camera provided on the moving body;
GPS that acquires image position information that is information of a position where the camera has captured an image;
A first storage unit for storing the image and the image position information;
A second accumulating unit for accumulating a target image feature quantity that is an image feature quantity of the target object and installation position information as information on the target object to be detected;
An image selection unit that selects the image as a selection image based on the image position information and the installation position information, and determines a designated region in the selection image as a detection region;
A detection unit that detects an object by determining whether or not the object exists with respect to the detection region using a similarity between the image feature amount extracted from the detection region and the target image feature amount; ,
And an output unit configured to notify a user when it is determined that the object does not exist. The object detection system according to claim 1,
When the detection unit determines that the object does not exist,
The image selection unit selects, as an auxiliary selection image, images positioned before and after the selection image from the first storage unit,
The said detection part performs an object detection with respect to the designated area | region in the said auxiliary | assistant selection image, The object detection system characterized by the above-mentioned. The object detection system according to claim 1,
A position estimation unit that estimates position information of the selected image as a detection position of the object when the detection unit determines that the object exists;
A determination unit that calculates a difference in distance between the detection position and the installation position information, and determines whether the difference is larger or smaller than a threshold; and
The object detection system according to claim 1, wherein when the determination unit determines that the difference is greater than a threshold value, the output unit notifies the user. The object detection system according to claim 3,
It further has a sensor unit that measures the distance to an obstacle located within the measurement range,
The position estimation unit is an object detection system that uses the distance measured by the sensor unit to estimate the detection position. The object detection system according to claim 1,
The camera has a first camera that images the traveling direction of the moving body, and a second camera that images the ground on which the moving body moves,
The second storage unit further stores the type of the object,
The image selection unit may be one of a first image captured by the first camera and a second image captured by the second camera based on the installation position information, the type, and the image position information. An object detection system, wherein the selected image is selected from an image. The object detection system according to claim 5,
An object detection system that switches a frame rate of the second camera when the object is detected from the first image. The object detection system according to claim 1,
When the detection unit determines that the object is present, a detected image that is an image of the object, an image feature amount extracted from the detected image, a time when the detected image is captured, and weather A third storage unit that stores status information that is information of
The detection unit performs object detection using an image feature amount stored in the third storage unit instead of the target image feature amount. The object detection system according to claim 7,
The determination unit calculates the amount of change by comparing the detected image with the detected image stored in the third storage unit in the past,
When the change amount is higher than a threshold value, the output unit notifies the user of the object detection system. The object detection system according to claim 7,
The object detection system further includes a DB selection unit that selects the detected image having the same situation information as the situation information of the selected image. A first step of storing, in a first storage unit, an image captured by a camera provided on a moving body and image position information that is information on a position where the camera has captured an image;
A second step of storing, in the second storage unit, target image feature quantities that are image feature quantities of the target objects and installation position information indicating positions where the target objects are installed as information on the target objects to be detected;
A third step of selecting the image as a selected image based on the image position information and the installation position information, and determining a designated area in the selected image as a detection area;
Fourth step of performing object detection by determining whether or not the object exists with respect to the detection region by using a similarity between the image feature amount extracted from the detection region and the target image feature amount. When,
And a step of notifying a user when it is determined that the object does not exist.

Images