TWI781760B - Application of deep learning image recognition technology to door collision detection method and system - Google Patents

Abstract

A method and system for applying deep learning image recognition technology to vehicle door collision detection, which includes an image capture device, an information processing unit, and an information output unit. The image capture device can continuously capture images of the rear of the car and form multiple dynamic images. The image recognition processing module of the information processing unit establishes a deep learning algorithm model, removes the image part outside the preset region of interest of each dynamic image, and uses the image to perform image feature extraction and then input it to the image recognition model to make the image recognition The model performs image recognition processing of object classification and object positioning on dynamic images to predict the category information and location information of moving objects, and judge whether it will enter the threshold range of the warning zone. If the judgment result is yes, it will output a warning signal. The display screen of the information output unit can display prediction frames and warning frames that include frame selection of moving objects to represent position information, so that by integrating deep learning image recognition and warning threshold technologies, drivers and passengers can obtain The movement of people and vehicles in the rear and the warning signal for the vehicles in the rear to enter the warning zone can greatly reduce the probability of being hit when the door is opened, thereby improving the safety protection of vehicles entering and exiting.

Description

Applying Deep Learning Image Recognition Technology to Vehicle Door Collision Detection Method and System

The present invention relates to a method and system for applying deep learning image recognition technology to car door collision detection, especially to a method and system that can enable drivers and passengers to obtain the dynamics of people and vehicles behind the car and issue entry warnings by integrating deep learning image recognition and warning threshold technologies Zone warning signal of car door opening anti-collision safety protection technology.

By the way, most of the doors of the car are pivotally connected with the car body, so each door can be independently opened and closed relative to the car body. Here, it is used as an entry and exit vehicle for passengers and drivers. the use of. However, you can often see some traffic accidents on the news or on social networking sites. The main reason is that the driver of the car did not open the door in two stages according to the regulations or did not pay attention to the car coming from behind and suddenly opened the door, which caused the rear locomotive to frighten and make the locomotive driver deviate. The line falls; or is suddenly hit by the open door and flies out of the line. No matter what kind of door opening event occurs, the locomotive driver will be injured seriously, and the locomotive driver will be seriously injured. In traffic accidents, this accident has indeed become common. Cause unnecessary casualties and property losses. Also because of this, government authorities in various countries have also faced up to the seriousness of this problem, and have begun to do relevant two-stage door opening operation safety enlightenment and propaganda; Types of traffic accidents have become a technical issue and challenge that governments, relevant academic circles, and industry players in various countries urgently want to solve.

In response to the above-mentioned shortcomings, related companies have developed a device that can be displayed before getting off the car and opening the door. The conventional technology of displaying the rear dynamic image, the representative patent of this conventional technology is shown in the patents such as CN1128600A "Vehicle Rear Confirmation Device" of Chinese Mainland Invention and No. M376460 "Door Anti-Collision Mirror" of the Republic of China New Announcement. These patents can close the door lock for a period of time before getting out of the car and opening the door, and display the rear dynamic image for the passenger or driver to watch, and then the passenger or driver decides whether to open the door. When a problem occurs and it is impossible to identify whether there is an oncoming car from the rear, a traffic accident in which the door is opened and bumped will still occur, thus causing great pressure, distress and inconvenience for passengers when the door is opened and get off.

Another known technology is to use sensing technology (such as ultrasonic sensor, infrared sensor and radar sensor) to sense the dynamics of people and vehicles approaching the rear of the vehicle. When a vehicle approaches, the door is temporarily closed. Locking and jamming cannot be opened, and when no vehicle is approaching, the door is unlocked and can be opened; representative patents of this known technology such as the Chinese Mainland New Model No. CN201633624 "Car Door Opening Safety Monitoring Device", the Republic of China New Announcement Patents such as No. 516534 "Door Opening Warning Device for Checking Cars Coming from Behind" and ROC New Announcement No. M318541 "Structure Improvement of Vehicle Door Warning Device" are shown; although these patents can use the above-mentioned sensing technology to sense the driving behind the vehicle dynamics; however, the distance between the sensing range of the ultrasonic sensor and the infrared sensor is not enough to capture the dynamics of a fast-moving vehicle, so that there will be problems with sensing dead angles and blind spots, so door openings will also be hit traffic accidents; in addition, although the sensing range of the radar is relatively long, the cost is too high, so that it cannot be fully popularized and adopted, thus causing an important obstacle to the realization of commercialization.

There is another such as the Republic of China Invention Announcement No. I531500 "Automobile Door Opening Warning Method Using Image Processing", the image processing unit of this patent can convert the obtained target object in the nth image according to the preset perspective conversion formula The pixel position of the center is converted to the The two-dimensional coordinates of the position of the target object on the road surface; and determine whether the distance between the target object and the car door of the car is less than a preset distance; when the determination result is yes, the image processing unit outputs a warning to the output unit. Although this patent has the function of image recognition of the distance between the target object and the car door to issue a warning; however, this patent does not integrate deep learning image recognition and the establishment of warning threshold range technology, so the distance value can only be used as the basis for warning release Moreover, the patent does not use whether the target object enters the warning threshold range as the basis for issuing the warning. The error is too large to cause false positives, thus causing inconvenience and troubles in use.

In view of this, the inventor believes that the above-mentioned prior art and the prior patents are indeed not perfect, and there is still a need for further improvement. Therefore, the inventors conducted further research and developed a patent as disclosed in the present invention. technical achievements.

The first object of the present invention is to provide a method and system for applying deep learning image recognition technology to vehicle door collision detection, mainly by integrating deep learning image recognition and warning threshold technology, so that drivers and passengers can obtain the dynamics and signals of people and vehicles behind the car. The warning signal of the rear vehicle entering the warning area can greatly reduce the probability of being hit when the door is opened, thereby improving the safety protection of car entry and exit. The technical means to achieve the first objective of the present invention includes an image capture device, an information processing unit and an information output unit. The image capture device can continuously capture images of the rear of the car and form multiple dynamic images. The image recognition processing module of the information processing unit establishes a deep learning algorithm model, and removes the image parts other than the preset interest area of each dynamic image In addition, image features are extracted by images and then input to the image recognition model, so that the image recognition model performs image recognition processing for object classification and object positioning on dynamic images, so as to predict the category information and location information of moving objects, and judge whether they will Enter the threshold range of the warning zone, if the judgment result is yes, then output a warning signal. The display screen of the information output unit can display a prediction frame and a warning frame that include frame selection of moving objects to represent position information.

The second object of the present invention is to provide a method and system for applying deep learning image recognition technology to vehicle door collision detection with the function of merging objects so that the prediction frame can quickly frame a more accurate range. The technical means to achieve the second objective of the present invention includes an image capture device, an information processing unit and an information output unit. The image capture device can continuously capture images of the rear of the car and form multiple dynamic images. The image recognition processing module of the information processing unit establishes a deep learning algorithm model, removes the image part outside the preset region of interest of each dynamic image, and uses the image to perform image feature extraction and then input it to the image recognition model to make the image recognition The model performs image recognition processing of object classification and object positioning on dynamic images to predict the category information and location information of moving objects, and judge whether it will enter the threshold range of the warning zone. If the judgment result is yes, it will output a warning signal. The display screen of the information output unit can display a prediction frame and a warning frame that include frame selection of moving objects to represent position information. Wherein, the image recognition processing module further includes an object merging step during execution, which includes the following steps: Step 1, sorting the predicted related categories from large to small according to the degree of confidence; Step 2, judging whether there are human objects, objects It is to frame other categories. If there is a human object, it will use IoU to judge whether there is any overlap between the human object and other predicted categories; step 3, when the IoU parameter is greater than the threshold range, Then merge the two objects into a new object; and step 4, if compared and merged When there is no repetition, frame the predicted category separately.

The third object of the present invention is to provide a door collision detection method and system that can issue an early warning against the danger caused by a small car dodging a large car and suddenly approaching the door. The technical means to achieve the third objective of the present invention includes an image capture device, an information processing unit and an information output unit. The image capture device can continuously capture images of the rear of the car and form multiple dynamic images. The image recognition processing module of the information processing unit establishes a deep learning algorithm model, removes the image part outside the preset region of interest of each dynamic image, and uses the image to perform image feature extraction and then input it to the image recognition model to make the image recognition The model performs image recognition processing of object classification and object positioning on dynamic images to predict the category information and location information of moving objects, and judge whether it will enter the threshold range of the warning zone. If the judgment result is yes, it will output a warning signal. The display screen of the information output unit can display a prediction frame and a warning frame that include frame selection of moving objects to represent position information. Wherein, when the image recognition processing module recognizes that the object is a bicycle and a large vehicle or a motorcycle and a large vehicle, and it is determined that the bicycle or the motorcycle judges that the object will not enter the warning zone and exceed the threshold range, Then send out the second warning signal that may enter the warning zone and exceed the threshold range.

1: car

10: Image capture device

20: Information processing unit

21: Image recognition processing module

210:Deep Learning Calculus Model

211: Coordinate parameter database

30: Information output unit

31: display screen

32:Warning device

40: Power supply unit

50:Bluetooth communication system

60: Memory module

A1: Region of interest

A2: warning area

A3: Threshold range

A4: Forecast grid

A5: warning sash

B1, B2: center pixel position

FIG. 1 is a schematic functional block diagram of a specific implementation architecture of the present invention.

FIG. 2 is a schematic diagram of the implementation process of the deep learning calculation module in the prediction stage of the present invention.

FIG. 3 is a schematic diagram of the implementation process of the deep learning algorithm module in the training phase of the present invention.

FIG. 4 is a schematic diagram of the implementation of displaying dynamic image classification and positioning information on the display screen of the present invention.

FIG. 5 is a schematic diagram showing the implementation of the display screen before combining two objects in the dynamic image according to the present invention.

FIG. 6 is a schematic diagram of the implementation of the combination of two objects in a dynamic image displayed on the display screen of the present invention.

FIG. 7 is a schematic diagram of another implementation of displaying dynamic image classification and positioning information on the display screen of the present invention.

FIG. 8 is a schematic diagram of the implementation of image processing to obtain the center pixel position of an object image according to the present invention.

FIG. 9 is a schematic diagram of the coordinate track analysis of the central pixel position of the image of the object in FIGS. 7 and 8 .

Fig. 10 is a schematic diagram of the process control implementation of the merged object of the present invention.

Fig. 11 is a schematic diagram of LOSS after the training of the deep learning calculation model of the present invention is completed.

FIG. 12 is a schematic diagram of the implementation of the weight test after the training of the deep learning calculation model of the present invention is completed.

In order to allow your examiner to further understand the overall technical characteristics of the present invention and the technical means to achieve the purpose of the present invention, the specific embodiments are described in detail in conjunction with the drawings: please refer to Figures 1 to 4, in order to achieve the first part of the present invention The first embodiment of an object includes an image capture device 10, an information processing unit 20 (such as a computer or a microprocessor MCU; but not limited thereto), an information output unit 30 for supplying required The power supply unit 40 of the power supply (such as a combination of a car battery and a power processing circuit; but not limited thereto). The image capture device 10 is set on the car 1 (such as the left rear-view mirror; or the trunk, etc.) and faces the rear of the car 1 with the camera lens, and can automatically or semi-automatically continuously monitor the people behind the car 1 or on the left side during the time of getting off the car. Car dynamic image capture and imaging into a plurality of continuous dynamic images. The information processing unit 20 includes an image recognition processing module 21, and establishes a deep learning algorithm model 210. The image recognition processing module 21 sets a preset region of each input dynamic image as a region of interest A1, And the part of the image other than the region of interest A1 is removed, so as to extract the image features of the dynamic image that retains the region of interest A1, and then input it to the deep learning calculation model 210, so that the deep learning calculation model 210 can perform a dynamic image analysis for each input Carry out image recognition processing of object classification and object positioning to predict the category information and location information of moving objects in the dynamic image, and judge whether the object will enter the threshold range of the default warning area A2 based on the category information and location information A3, when the judgment result is yes, then output a warning signal. The information output unit 30 includes a display screen 31 for displaying dynamic images And a warning device 32 for outputting the warning signal as audio or voice warning information (such as the danger that the door cannot be opened, and there will be a vehicle passing behind the car). Prediction frame A4 representing location information, category information (such as automobile, motorcycle, bicycle or person) near prediction frame A4 and warning frame A5 representing warning area A2, if the object in prediction frame A4 will pass the threshold In the range A3, the prediction box A4 will give a warning with a red box and flashing, so that the driver or passengers can know that they are in a dangerous state at this time and cannot open the door to get off; if the object in the prediction box A4 will not pass the threshold The range A3 and the prediction frame A4 will be represented by a green frame, so that the driver or passengers can know that the processing is safe at this time, and you can safely open the door and get off the car.

Please refer to the embodiment shown in Figures 2 to 3, the image recognition processing module 21 includes the following steps when executed:

(a) The step of the training stage is to establish a deep learning calculation model 210, input the feature samples of a huge number of objects, warning area identification parameters, image classification parameters and image positioning parameters into the deep learning calculation model 210, and test by the deep learning calculation model 210 The correct rate of image recognition is then judged whether the correct rate of image recognition is sufficient. If the judgment result is yes, then the recognition result is output and stored.

(b) The step of running the prediction stage is to input the real-time captured dynamic images into the deep learning algorithm model 210 in sequence, and calculate the corresponding object features by the deep learning algorithm model 210, so as to predict the category information and location information of the recognized object and identification result information such as whether the object exceeds the threshold range A3 preset by the alert area A2.

In a specific embodiment of the present invention, the deep learning calculation model 210 uses YOLO v4 as the network architecture to detect object classification and location identification, and the data set of the deep learning calculation model 210 uses the VOC data set. Using the VOC2012 data set as training The stage step is used, and the VOC2007 data set is used as the prediction stage step.

Please refer to Figures 1 to 6 and Figure 10, in order to achieve the second embodiment of the second purpose of the present invention, this embodiment is based on the image recognition processing module 21 in addition to the overall technical content of the first embodiment above. Execution further includes an object merging step, which includes the following steps:

Step 1: sort the predicted related categories in descending order of confidence.

Step 2. Determine whether there is a human object, and the object is to frame other categories. If there is a human object, it will judge whether there is an overlap between the human object and other predicted categories in the form of IoU.

Step 3, when the IoU parameter is greater than the threshold range A3, merge the two objects into a new object, as shown in Figures 5 and 6, merging the motorcycle and the person into a knight.

Step 4, if there is no duplication after comparison, frame the predicted category respectively.

1, 7 and 8, the image recognition processing module 21 establishes a coordinate parameter database 211, and the coordinate parameter database 211 is set to store a coordinate parameter data corresponding to the dynamic image and a frame for surrounding To locate the threshold parameter data of the warning area A2, the image recognition processing module 21 further includes an image positioning step, which is to convert the dynamic image retaining the area of interest A1 from the RGB format to the YIQ plane, and then obtain it from the YIQ plane Object image, and then use the I value of the YIQ plane, take the threshold to remove most of the image, keep only the object image of the object, and then filter out the small area of non-object image through small area noise filtering to obtain the bottom or left bottom of the object image The center pixel position, and then bring the center pixel position and threshold parameter data into the coordinate parameter data The material warehouse 211 calculates the actual coordinate positions of the center pixel positions B1, B2 and the threshold range A3 of the bottom of the object image (as shown in Figure 7, the bottom of the locomotive tire); or the left bottom (as shown in Figure 7, the bottom of the left tire of the bus), respectively. It is used as the basis for judging that the object image enters the warning area A2 and reaches the threshold range A3.

Specifically, please refer to Figure 8 again. The central pixel position B1 will pass through the threshold range A3, so a warning will be issued immediately. As for the central pixel position B2, although it will pass through the warning area A2, it will not pass through the threshold range A3, so it will not Issue a warning.

In a specific embodiment of the present invention, the information output unit 30 can be one of smart phones, smart watches, smart bracelets and smart glasses, and the information output unit 30 communicates with the information processing unit 20 through the Bluetooth communication system 50 Information links, and use its display screen 31 to display dynamic images, category information and location information, and then use its warning device 32 to output warning information.

Please refer to Figures 1 to 4 and Figure 7, the third embodiment to achieve the third purpose of the present invention, this embodiment includes the overall technical content of the first embodiment, when the image recognition processing module 21 image recognition When the second object is a bicycle and a large vehicle; or a motorcycle and a large vehicle parallel and it is determined that the bicycle or motorcycle will not enter the warning zone A2 and the threshold range A3, then a second warning signal that may enter the warning zone A2 and the threshold range A3 will be issued , let the driver and passengers know this parallel situation, so that the dangerous situation caused by dodging the large vehicle and approaching the direction of the car door of the car suddenly can be avoided.

In the application embodiment of the present invention, it is a driving recorder, as shown in Figure 1, it includes a memory module 60 for recording continuous dynamic dynamic images, an image capture device 10, and an information processing unit 20 , memory module 60 and warning device 32 accommodating casing groups (this figure formula example is not shown), the casing group is located in the car 1 near the rear window so that the camera lens can face the position behind the car.

It must be stated that the time period for getting off the car mentioned above can be realized by capturing the original door opening signal and door closing signal of each door switch of the car 1 to trigger the information processing unit 20; When it is opened to the first angle, the door switch of the car door will generate a door-opening signal, and when the information processing unit 20 is triggered by the door-opening signal, it will be identified as the start time of the getting-off period, and the image capture device 10 will be activated to carry out Image capture; when the information processing unit 20 is triggered by the door closing signal, it will be determined as the end time of the getting off period, and the image capture device 10 will be closed to end the image capture operation.

The above-mentioned region of interest A1 (Region of Interest, RoI) is to design a region of interest A1 for dynamic images, as shown in Figure 4. A region is framed from the image, and the region selected by the frame will be a part of the analysis. One of the key points is to carry out the next step of processing on the area selected by the frame, so a larger image to be processed will become a small area of the image, which can greatly reduce the processing time and increase the accuracy.

Figure 11 shows the relevant LOSS value performance of YOLOv4 after training from VOC2012. From this figure, it can be seen that the LOSS can be reduced very quickly and the convergence is completed during the training process. The related val loss is also similar to the train loss. Figure 12 shows the test of the weights completed in real-time training. The test data is the VOC2007 data set. It can be seen that the person in the figure reaches 85% AP, while the car and motorbike have nearly 80% AP, but the accuracy on the bicycle Approaching 60% AP, and the four categories have a mAP of 74.76% in the case of IoU 0.5 performance, so in untrained pictures, the performance of YOLOv4 is still very prominent, and through the test data, we can also know that the weight has not caused the problem of over-fitting (Over-fitting).

YOLOv4 network architecture can be divided into input (Input), skeleton (Backbone), neck (Neck), head (Head), and the head can be divided into Dense Prediction (one stage) and Sparse Prediction (two stages), YOLOv4 The skeleton uses CSPDarknet53, the neck uses SPP and PAN, and the head uses YOLOv3. YOLOv4 is outstanding in FPS performance and AP performance. Comparing YOLOv4 with ASFF, it can be seen that YOLOv4's AP accuracy is close to FPS70, and its AP performance is close to 43.5%, while ASFF's FPS is close to 30. 40%, which is far from the performance of YOLOv4. Compared with the EfficienDet network, although the AP of YOLOv4 and EfficienDet is nearly as high as 50%, the FPS performance is only about 12. There is still a distance from real-time computing. The performance AP of the calculation is lower than 43.5% of YOLOv4. Therefore, the present invention uses the network architecture of YOLOv4. The AP is the average value of the predicted result and the benchmark truth greater than the IoU threshold 0.50:0.05:0.95. All 10 IoU thresholds are all 80 average of the categories.

Furthermore, when the YOLOv4 network outputs, it will predict people, locomotives, bicycles, and car objects respectively. The prediction results are shown in Figure 5. The set overlapping degree will be merged later to form a new object called rider (rider). After the merge, it will be displayed as shown in Figure 6.

As shown in Figure 4, when the rider does not enter the threshold range A3 of the warning zone A2, No Alarm will be displayed above the center of the image. When the knight enters the warning zone A2, a warning will be displayed. Thresholds in zone A2 will not display warnings.

The present invention adopts the deep learning YOLOv4 to detect the coming car system. When the car 1 forgets to open the door in two stages or does not pay attention to the car coming from behind, an additional auxiliary tool warns the car 1 whether there is a car coming from behind or is warning The beep warns driving, the architecture of YOLOv4 can quickly identify objects and can accurately frame the position and object category, and has a recognition speed of nearly 15 (Fps; frames per second) in the input size of 608x608. In the collision detection, the judgment ability of even detection can be achieved.

After the above detailed description, the present invention has the following features:

1. The present invention can indeed, by integrating technologies such as deep learning image recognition and warning threshold range, allow drivers and passengers to obtain the dynamics of people and vehicles behind the car and issue a warning signal of entering the warning area, so as to greatly reduce the impact on the door opening The probability of occurrence, thereby improving the safety protection of car access.

2. The present invention is indeed a deep learning image recognition technology that has the function of merging objects so that the prediction frame can frame an accurate range more quickly.

3. The present invention is indeed an application of deep learning image recognition technology that can issue a warning against the danger caused by the small car dodging the large car and suddenly approaching the door.

The above is only a feasible embodiment of the present invention, and is not intended to limit the patent scope of the present invention. All equivalent implementations of other changes based on the content, features and spirit of the following patent scopes are cited. All should be included within the scope of the present invention. The structural features of this invention, which are defined in the scope of the patent application, are not found in similar products, and are practical and progressive, and have met the requirements of an invention patent. Please file an application in accordance with the law. I would like to ask the Jun Bureau to approve the patent in accordance with the law. Safeguard the legitimate rights and interests of the applicant.

10: Image capture device

20: Information processing unit

21: Image recognition processing module

211: Coordinate parameter database

30: Information output unit

31: display screen

32:Warning device

40: Power supply unit

50:Bluetooth communication system

60: Memory module

Claims (10)

A method of applying deep learning image recognition technology to vehicle door collision detection, comprising: Provide an image capture device, an information processing unit, and an information output unit; wherein, the information processing unit includes an image recognition processing module, and the information output unit includes a display screen and a warning device; The image capture device is installed on a car with a lens facing the rear of the car, and is used to continuously capture images of the dynamics of people and vehicles at the rear during the time of getting off the car and form multiple dynamic images; A deep learning algorithm model is established in the image recognition processing module, and the image recognition processing module sets a preset region of each input dynamic image as a region of interest, and sets the image part outside the region of interest Remove, to extract the image features of the dynamic image that retains the region of interest and then input it to the deep learning calculation model, so that the deep learning calculation model can classify the object and the object for each input dynamic image Positioning image recognition processing to predict the type information and location information of the object moving in the dynamic image, and judge whether the object will enter a threshold range of a warning zone based on the type information and the location information, and judge the result Yes, output a warning signal; and Use the display screen to display the dynamic image, and use the warning device to output the warning signal as audio or voice warning information, and then display the moving object on the display screen including the object in frame selection to represent the At least one prediction frame of location information, the category information located near the at least one prediction frame, and a warning frame representing the warning area. As described in Claim 1, the application of deep learning image recognition technology to the vehicle door collision detection method, wherein the image recognition processing module includes the following steps when executed: (a) The step of the training stage is to establish the deep learning calculation model, and input a huge amount of feature samples of the object, warning area identification parameters, image classification parameters and images into the deep learning calculation model Positioning parameters, and the deep learning calculation model tests the accuracy of image recognition, and then judges whether the accuracy of image recognition is sufficient. If the judgment result is yes, the recognition result is output and stored; when the judgment result is no, the depth learning algorithm model self-modification learning; and (b) The step of running the prediction stage is to sequentially input the dynamic image captured in real time into the deep learning calculation model, and calculate the corresponding characteristics of the object by the deep learning calculation model, so as to predict and recognize the category of the object information, the location information, and identification result information such as whether the object exceeds the preset threshold range of the warning zone. Applying deep learning image recognition technology to the car door collision detection method as described in claim 1, wherein the deep learning algorithm model is to use YOLO v4 as the network architecture for detection of object classification and positioning recognition, the deep learning algorithm The data set of the model uses the VOC data set, the VOC2012 data set is used as the training stage step, and the VOC2007 data set is used as the prediction stage step. The application of deep learning image recognition technology to the car door collision detection method as described in claim 3, wherein the VOC data set is trained, verified and processed on the deep learning algorithm model for the four categories of people, cars, motorcycles and bicycles. test. As described in Claim 1, the application of deep learning image recognition technology to the vehicle door collision detection method, wherein the image recognition processing module further includes an object merging step during execution, which includes the following steps: Step 1, sort the predicted related categories in order of confidence from large to small; Step 2. Determine whether there is a human object. The object is to frame other categories. If there is a human object, it will be judged by IoU against the human object and other predicted categories. whether there are overlapping parts; Step 3, when the parameter from IoU is greater than the threshold range, merge the two objects into a new object; and Step 4, if there is no duplication after comparison, frame the predicted category respectively. As described in claim 1, the application of deep learning image recognition technology to the vehicle door collision detection method, wherein the image recognition processing module establishes a coordinate parameter database, and the coordinate parameter database is set to store a coordinate parameter corresponding to the dynamic image Data and a threshold parameter data used to surround and locate the warning area, the image recognition processing module further includes performing an image positioning step, which is to convert the dynamic image with the region of interest from RGB format to YIQ plane, then obtain the image of the object from the YIQ plane, then use the I value of the YIQ plane, take the threshold value to remove most of the image, and only keep the object image of the object, and then filter out the small area of the non-target image by small-area noise filtering Object image to obtain the center pixel position of the bottom or left bottom of the object image, and then bring the center pixel position and the threshold parameter data into the coordinate parameter database to calculate the center pixel position and the threshold of the object image respectively The coordinate position of the range is used as the basis for judging that the object image enters the warning zone and exceeds the threshold range. The application of deep learning image recognition technology to the vehicle door collision detection method as described in claim 1, wherein, when the image recognition processing module recognizes that the object is a bicycle and a large vehicle or a locomotive and a large vehicle, it is determined that the bicycle Or when the locomotive will not enter the warning zone and the threshold range, it will send a second warning signal that it may enter the warning zone and exceed the threshold range. A vehicle door collision detection system applying deep learning image recognition technology, comprising: An image capture device, which is installed on a car and faces the rear of the car with a lens, for Continuously capture images of the dynamics of people and vehicles behind the vehicle during the time of getting off the vehicle to form multiple dynamic images; An information processing unit, which includes an image recognition processing module, and establishes a deep learning algorithm model, and the image recognition processing module sets a preset area of each input dynamic image as an interest area, and The part of the image outside the region of interest is removed, so as to extract the image features of the dynamic image that retains the region of interest, and then input it to the deep learning algorithm model, so that the deep learning algorithm model can be used for each input of the dynamic image. Perform object classification and image recognition processing for the location of the object to predict the category information and location information of the object moving in the dynamic image, and judge whether the object will enter a warning based on the category information and the location information One of the threshold ranges of the zone, if the judgment result is yes, then output a warning signal; and An information output unit, which includes a display screen for displaying the dynamic image and a scene display device for outputting the warning signal as audio or voice warning information, the dynamic image displayed on the display screen includes a frame selection The moving object is represented by at least one predicted frame of the location information, the category information near the at least one predicted frame, and a warning frame representing the warning area. Applying deep learning image recognition technology to the car door collision detection system as described in claim 8, wherein the information output unit is selected from one of smart phones, smart watches, smart bracelets and smart glasses, and the information output unit is passed through The bluetooth communication system is connected with the information processing unit, and uses its display screen to display the dynamic image, the category information and the location information, and then uses its warning device to output warning information. A driving recorder applying the deep learning image recognition technology to the door collision detection system of application claim 8, which includes a memory module for recording the dynamic image of continuous dynamics and A housing set for the image capture device, the information processing unit, the memory module and the warning device, the housing set is installed in the car near the rear window so that the lens can face the rear .

Images