Method for judging head orientation of pedestrian falling to ground after pedestrian-vehicle collision

Abstract

The invention discloses a pedestrian landing head orientation determination method after a pedestrian-vehicle collision, and the method comprises the steps: obtaining the image data of the exterior of a vehicle when the pedestrian-vehicle collision occurs, and obtaining a key frame containing the pedestrian information through image segmentation; performing image processing on the key frame, and extracting a human body region and a candidate frame in the image; extracting attitude key points of the pedestrian from the key frame by using an Alphapose algorithm; constructing a human skeleton model according to the attitude key points, analyzing coordinates of five key points of the head of the pedestrian in the landing process based on the human skeleton model, and calculating a vector relationship between the key points by converting coordinate information of the pedestrian key points in an image between two key time nodes of collision and landing into head motion information; determining the head landing orientation according to the change of the vector relationship; according to the invention, the specific orientation of the head of the pedestrian can be judged under the condition of poor visibility, and a medical institution can be helped to carry out preliminary diagnosis on pedestrian injury after a traffic accident of the pedestrian occurs.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine vision recognition, and in particular relates to a method for determining the direction of a pedestrian's head when the pedestrian falls to the ground after a collision between a pedestrian and a vehicle. Background Art

[0002] Pedestrians are in a vulnerable position in road traffic accidents. After a collision between a pedestrian and a vehicle, the main injuries to the pedestrians are the head, upper limbs, chest, and spine. Among them, head injuries occur most frequently. The treatment of head injuries is closely related to time. If the treatment is not timely, it will threaten the life safety of pedestrians.

[0003] The existing Advanced Automatic Crash Notification System (AACN) effectively protects people inside the car but ignores the protection of people outside the car. This is because after a collision between a pedestrian and a car, when the pedestrian falls to the ground, the clarity of the surveillance video is poor and it is impossible to determine the direction of the pedestrian's head, which in turn affects the protection effect of AACN on people outside the car. Summary of the invention

[0004] In order to address the deficiencies in the prior art, the present invention proposes a method for determining the direction of a pedestrian's head when landing after a collision between a person and a vehicle. The method establishes a human skeleton structure model based on the detection of human key points using Alphapose, obtains the key points of the head, converts the coordinate information of the pedestrian's key points in the image between the two key time nodes of impact and landing into head movement information, calculates the vector relationship between the key points, and determines the direction of the head landing according to the change in the vector relationship.

[0005] The present invention achieves the above technical objectives through the following technical means.

[0006] A method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle comprises the following steps:

[0007] Step 1: Obtain image data of the vehicle exterior when a collision occurs, and obtain key frames containing pedestrian information through image segmentation;

[0008] Step 2: Perform image processing on the key frame to extract the human body area and candidate frame in the image;

[0009] Step 3: Use the Alphapose algorithm to extract the pedestrian's posture key points from the key frames;

[0010] Step 4: Based on the posture key points obtained in step 3, a human skeleton model is constructed. Based on the human skeleton model, the coordinates of the five key points of the pedestrian's head during the landing process are analyzed, and the judgment result of the pedestrian's head direction when landing is output.

[0011] Furthermore, the process of judging the direction of the pedestrian's head when landing is as follows:

[0012] Step 4.1: Convert the coordinate information of the posture key points in the key frame into the pixel coordinate information of the corresponding frame;

[0013] Step 4.2: Establish human skeleton models in different frames according to the corresponding pixel coordinate information;

[0014] Step 4.3: Determine the direction of the pedestrian's head when landing based on the time difference between key frames and the vector relationship between the posture key points of the human skeleton model between each frame.

[0015] Furthermore, the top left corner vertex of the key frame image is used as the coordinate origin, and the row index and column index in the pixel coordinate system are used as the X-axis and Y-axis, respectively, to establish a pixel coordinate system. The five key points of the head are represented as follows: The key point P representing the nose 0 The corresponding coordinates can be expressed as (x 0 ,y 0 ), represents the left eye key point P 1 It is expressed as (x 1 ,y 1 ), represents the key point P of the right eye 2 It is expressed as (x 2 ,y 2 ), represents the left ear key point P 3 It is expressed as (x 3 ,y 3 ), represents the key point P of the right ear 4 It is expressed as (x 4 ,y 4 ).

[0016] Furthermore, the pedestrian's head direction when landing is divided into landing with the front of the head, landing with the back of the head, landing with the left side of the head and landing with the right side of the head.

[0017] Furthermore, the basis for judging the direction of the pedestrian's head when landing based on the five key points of the head is:

[0018] (1) Judgment basis for head frontal landing: Extract the coordinate information of the five key points of the pedestrian's head during landing. If the relative positions of the five key points of the head remain unchanged and P 0 and P 1 The horizontal coordinates of x are consistent, that is, 1 =x 2 ; The key point representing the nose and the key point representing the left eye P at the moment when the pedestrian's head lands 0 ' and P 1 The absolute value of the difference between the vertical coordinates of

[0019] |y 1 '-y0 '| is less than the previous moment P 0 and P 1 The absolute value of the difference between the ordinates of |y 1 -y 0 |If the above conditions are met, it is considered that the pedestrian’s head has landed directly on the ground.

[0020] (2) The right side of the head lands on the ground: The coordinate information of the five key points of the pedestrian’s head is extracted. If the current moment P 0 and P 1 The absolute value of the difference between the ordinates of |y 1 -y 0 |Than the next moment P 0 ' and P 1 'The absolute value of the difference between the vertical coordinates |y 1 '-y 0 '| Large case, while P O and P1 composed of vectors and P 0 ' and P 1 'When the angle between the composed vectors gradually decreases and the result is finally calculated to be within the range of 0°-45°, it is determined that the right side of the head has landed.

[0021] (3) The left side of the head lands: Extract the coordinate information of the five key points of the pedestrian’s head during the landing process. If the current moment P 0 and P 1 The absolute value of the difference between the ordinates of |y 1 -y 0 |Equal to P at the next moment 0 ' and P 1 'The absolute value of the difference between the vertical coordinates |y 1 '-y 0 '|, and at the same time by P O and P 1 The vector and P 0 and P 2 The angle between the vectors is 0°, and P 0 ,P 1 ,P 2 The horizontal axis is consistent, P 0 The vertical coordinate is greater than P 1 ,P 2 When the vertical coordinate is 0, it is judged that the left side of the head lands.

[0022] (4) The back of the head touches the ground: Extract the coordinate information of the five key points of the pedestrian’s head during the landing process. If the current moment P 0 and P 1 The absolute value of the difference between the ordinates of |y 1 -y 0 |Equal to P at the next moment0 ' and P 1 'The absolute value of the difference between the vertical coordinates |y 1 '-y 0 '|, and at the same time by P 0 and P 1 The vector composed of 0 and P 2 The angle between the vectors is 0°, and the coordinate point P 0 , P 1 and P 3 The vertical coordinates are equal and P 0 , P 1 If they overlap, it is determined that the back of the head has landed.

[0023] Furthermore, the process of extracting the key points of the pedestrian's posture from the key frame in step 3 is as follows:

[0024] Step 3.1: Based on the obtained candidate frame, use the single person pose estimator (SPPE) in the Alphapose algorithm to perform in-depth analysis on the human body area in the candidate frame to obtain the key human body area;

[0025] Step 3.2: Use the spatial transformer network (STN) to analyze key human body regions;

[0026] Step 3.3: Use convolutional neural network to extract features of key human body areas;

[0027] Step 3.4: Use SPPE to process the extracted features to obtain the posture key points in the key human body area, use the posture guided region generator (PGPG) to process the posture key points, and correct the posture key points with large deviations;

[0028] Step 3.5: Use the reverse spatial variation network (SDTN) to map the pose keypoints into keyframes according to body parts.

[0029] Furthermore, the process of extracting the human body area and candidate frame in the image in step 2 is as follows:

[0030] Step 2.1: Take the key frame obtained in step 1 as the input image, transform the input key frame image, and locate the pedestrian in the image at the center of the detection frame;

[0031] Step 2.2: Use the region proposal network (RPN) to identify the human body region in all key frames, and the region proposal network (RPN) outputs the candidate box;

[0032] Step 2.3: Use non-maximum suppression to perform confidence detection on the candidate boxes to eliminate redundant candidate boxes.

[0033] Furthermore, the input key frame image is transformed using the spatial transformer network (STN).

[0034] Furthermore, in step 2.3, non-maximum suppression (NMS) is used to perform confidence detection on the candidate box, and non-maximum suppression includes confidence elimination and distance elimination.

[0035] Furthermore, step 1 uses the YoloV5-CBAM algorithm to extract key frames containing pedestrian information from the acquired vehicle external image data.

[0036] Beneficial effects of the present invention:

[0037] The present application proposes a method for determining the orientation of a pedestrian's head when the pedestrian falls to the ground after a collision between a person and a vehicle. Based on the Alphapose algorithm, a schematic diagram of key points of the human body is established to extract the coordinates of key points of the head. Finally, the specific orientation of the head is determined based on the coordinate difference and vector relationship between the key points obtained over time. The method can determine the specific orientation of the pedestrian's head under conditions of poor visibility, which is helpful for medical institutions to make a preliminary diagnosis of pedestrian injuries after a traffic accident. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 The present invention is a logical framework diagram of a method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle.

[0039] Figure 2 This is the detection flow chart of Alphaposee pedestrian posture.

[0040] Figure 3 Schematic diagram of key points of the human body.

[0041] Figure 4 Output image for Alphapose model result.

[0042] Figure 5 These are the four conditions in which a pedestrian's head lands on the ground. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0044] like Figure 1 The method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle is shown, comprising the following steps:

[0045] Step 1: Obtain image data of the vehicle exterior when a collision occurs, and obtain key frames containing pedestrian information through image segmentation. The details are as follows:

[0046] The YoloV5-CBAM algorithm is used to extract key frames containing pedestrian information from the acquired vehicle external image data, and images containing pedestrian information in the process from the collision between the vehicle and the pedestrian to the landing are extracted as key frames.

[0047] Step 2: Perform image processing on the key frame to extract the human body area and candidate frame in the image; Figure 2 As shown, the specific process is as follows:

[0048] Step 2.1: Take the key frame obtained in step 1 as the input image, use the spatial transformer network (STN) to transform the input key frame image, and locate the pedestrian in the image to the center of the detection box;

[0049] Step 2.2: Use the region proposal network (RPN) to identify the human body region in all key frames, and the region proposal network (RPN) outputs the candidate box;

[0050] Step 2.3: Use non-maximum suppression to perform confidence detection on the candidate boxes to eliminate redundant candidate boxes.

[0051] Furthermore, in step 2.3, non-maximum suppression (NMS) is used to perform confidence detection on the candidate box. Non-maximum suppression includes confidence elimination and distance elimination. The process is as follows:

[0052] Step 2.3.1: Confidence Elimination

[0053] Sort all detected candidate boxes according to their confidence, that is, for all candidate boxes {b 1 ,b 2 ,……b n}, according to its confidence score {C 1 ,C 2 ,……C n}Sort in descending order, select the frame with the highest score as the reference frame, and proceed to the next step of distance elimination.

[0054] Step 2.3.2: Distance Elimination

[0055] Use the intersection over union (IoU) to determine the degree of overlap between boxes and decide whether to suppress other boxes. The specific steps are as follows:

[0056]

[0057] Among them, bi represents the reference box with the highest confidence score in confidence elimination, bj represents other detection boxes except bi, and Area of ​​Intersection is box b i and box b j The intersection area of ​​​​the box bi and box b j The union area of ​​.

[0058] If the calculated IoU(b i ,b j )>T, T is the set detection threshold, then eliminate box b j , and repeat the calculation until all detection boxes are processed.

[0059] Step 3: Use the Alphapose algorithm to extract the pedestrian's posture key points from the key frames.

[0060] Step 3.1: Based on the obtained candidate frame, use the single person pose estimator (SPPE) in the Alphapose algorithm to perform in-depth analysis on the human body area in the candidate frame to obtain the key human body area, such as Figure 4 As shown in the figure, since SPPE is a single-person pose estimator, it can only identify a single human body region at a time. If there are multiple human body regions in the key frame, SPPE needs to be used multiple times to perform in-depth analysis of the human body region.

[0061] Step 3.2: Use the spatial transformer network (STN) to analyze key human body areas to ensure that SPPE is robust when performing posture estimation at different angles

[0062] Step 3.3: Use convolutional neural network to extract features of key human body areas;

[0063] Step 3.4: Use SPPE to process the extracted features to obtain the posture key points in the key human body area, such as Figure 3 As shown in the figure, the posture key points are processed using the posture guidance region generator (PGPG), and the posture key points with large deviations are corrected;

[0064] Step 3.5: Use the reverse spatial variation network (SDTN) to map the pose keypoints into keyframes according to body parts.

[0065] Step 4: Based on the posture key points obtained in step 3, a human skeleton model is constructed, and the direction of the pedestrian's head when landing is determined based on the human skeleton model.

[0066] Step 4.1: Based on the key frame with posture key point information obtained in step 3.5, the coordinate information of the posture key points in the key frame is converted into the pixel point coordinate information of the corresponding frame;

[0067] Step 4.2: Establish human skeleton models in different frames according to the corresponding pixel coordinate information;

[0068] Step 4.3: Determine the direction of the pedestrian's head when landing based on the time difference between key frames and the vector relationship between the posture key points of the human skeleton model between each frame.

[0069] Since the present invention is aimed at judging the direction of the pedestrian's head when landing, in general, the direction of the head landing and the corresponding injuries can be divided into the following four categories, such as Figure 5 As shown:

[0070] (1) Head-on impact: usually causes injuries related to the front of the head.

[0071] (2) Landing on the back of the head: This may involve injury to the back of the head.

[0072] (3) Landing on the left side of the head: This may cause damage to the left half of the head, and the temporal bone, temporal muscles and other structures may be involved.

[0073] (4) Landing on the right side of the head: Similar to landing on the left side, but it affects the right half of the head structure.

[0074] Therefore, this application extracts five key points located on the head from the posture key points, namely the key point P representing the nose 0 , left eye key point P 1 , right eye key point P 2 , left ear key point P 3 , right ear key point P 4 .

[0075] The present invention determines the direction of the pedestrian's head when landing based on the above five head key points. For the convenience of description, the upper left corner vertex of the key frame image is used as the coordinate origin, and the row index and column index in the pixel coordinate system are used as the X-axis and Y-axis respectively to establish a pixel coordinate system; therefore, in the key frame image, the corresponding position of each key point in the pixel coordinate system is determined, and the coordinates of each key point can be expressed as: the key point P representing the nose 0 The corresponding coordinates can be expressed as (x 0 ,y 0 ), represents the left eye key point P 1 It is expressed as (x 1 ,y 1 ), represents the key point P of the right eye 2 It is expressed as (x 2 ,y 2 ), represents the left ear key point P 3 It is expressed as (x 3 ,y 3 ), represents the key point P of the right ear 4 It is expressed as (x 4 ,y 4 ).

[0076] The basis for judging the direction of the pedestrian's head when landing based on the five key points of the head in the present invention is:

[0077] (1) Judgment basis for head frontal landing: Extract the coordinate information of the five key points of the pedestrian's head during landing. If the relative positions of the five key points of the head remain unchanged and P 0 and P 1 The horizontal coordinates of x are consistent, that is, 1 =x 2 ; The key point representing the nose and the key point representing the left eye P at the moment when the pedestrian's head lands 0 ' and P 1 The absolute value of the difference between the vertical coordinates of

[0078] |y 1 '-y 0 '| is less than the previous moment P 0 and P 1 The absolute value of the difference between the ordinates of |y 1 -y 0 |If the above conditions are met, it is considered that the pedestrian’s head has landed directly on the ground.

[0079] (2) The right side of the head lands on the ground: The coordinate information of the five key points of the pedestrian’s head is extracted. If the current moment P 0 and P 1 The absolute value of the difference between the ordinates of |y 1 -y 0 |Than the next moment P 0 ' and P 1 'The absolute value of the difference between the vertical coordinates |y 1 '-y 0 '| Large case, while P O and P1 composed of vectors and P 0 ' and P 1 'When the angle between the composed vectors gradually decreases and the result is finally calculated to be within the range of 0°-45°, it is determined that the right side of the head has landed.

[0080] (3) The left side of the head lands: Extract the coordinate information of the five key points of the pedestrian’s head during the landing process. If the current moment P 0 and P 1 The absolute value of the difference between the ordinates of |y 1 -y 0 |Equal to P at the next moment 0 ' and P 1 'The absolute value of the difference between the vertical coordinates |y 1 '-y 0 '|, and at the same time by P O and P 1 The vector and P0 and P 2 The angle between the vectors is 0°, and P 0 ,P 1 ,P 2 The horizontal axis is consistent, P 0 The vertical coordinate is greater than P 1 ,P 2 When the vertical coordinate is 0, it is judged that the left side of the head lands.

[0081] (4) The back of the head touches the ground: Extract the coordinate information of the five key points of the pedestrian’s head during the landing process. If the current moment P 0 and P 1 The absolute value of the difference between the ordinates of |y 1 -y 0 |Equal to P at the next moment 0 ' and P 1 'The absolute value of the difference between the vertical coordinates |y 1 '-y 0 '|, and at the same time by P 0 and P 1 The vector composed of 0 and P 2 The angle between the vectors is 0°, and the coordinate point P 0 , P 1 and P 3 The vertical coordinates are equal and P 0 , P 1 If they overlap, it is determined that the back of the head has landed.

[0082] In summary, the specific process of determining the head orientation of a pedestrian after landing is as follows: first, read the video of the entire process of the pedestrian landing and segment the video according to time. Secondly, use the Alphapose posture estimation algorithm to segment it into multiple images for processing, and save the position information of the key coordinate points of each image to generate a human skeleton model. Thirdly, select the key points of the head and calculate the moving distance and their vector angles between the key points based on the time difference, and use four pedestrian landing collision location algorithms to determine the final orientation of the pedestrian's head landing.

[0083] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle, characterized in that: The steps include: Step 1: Obtain image data of the vehicle exterior when a collision occurs, and obtain key frames containing pedestrian information through image segmentation; Step 2: Perform image processing on the key frame to extract the human body area and candidate frame in the image; Step 3: Use the Alphapose algorithm to extract the pedestrian's posture key points from the key frames; Step 4: Based on the posture key points obtained in step 3, a human skeleton model is constructed. Based on the human skeleton model, the coordinates of the five key points of the pedestrian's head during the landing process are analyzed, and the judgment result of the pedestrian's head direction when landing is output.

2. A method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle according to claim 1, characterized in that: The process of judging the direction of the pedestrian's head when landing is as follows: Step 4.1: Convert the coordinate information of the posture key points in the key frame into the pixel coordinate information of the corresponding frame; Step 4.2: Establish human skeleton models in different frames according to the corresponding pixel coordinate information; Step 4.3: Determine the direction of the pedestrian's head when landing based on the time difference between key frames and the vector relationship between the posture key points of the human skeleton model between each frame.

3. The method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle according to claim 2, characterized in that: Taking the upper left corner vertex of the key frame image as the coordinate origin, and taking the row index and column index in the pixel coordinate system as the X-axis and Y-axis respectively, a pixel coordinate system is established. The five key points of the head are represented as follows: the coordinates corresponding to the key point P0 representing the nose can be expressed as (x0, y0), the key point P1 representing the left eye is expressed as (x1, y1), the key point P2 representing the right eye is expressed as (x2, y2), the key point P3 representing the left ear is expressed as (x3, y3), and the key point P4 representing the right ear is expressed as (x4, y4).

4. A method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle according to claim 3, characterized in that: The direction of the pedestrian's head when landing is divided into landing with the front of the head, landing with the back of the head, landing with the left side of the head and landing with the right side of the head.

5. The method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle according to claim 4, characterized in that: The basis for judging the direction of the pedestrian's head when landing based on the five key points of the head is: (1) Judgment basis for head landing: extract the coordinate information of the five key points of the pedestrian's head during landing. If the relative positions of the five key points of the head remain unchanged, and the horizontal coordinates of P0 and P1 are consistent, that is, x1 = x2; the key point representing the nose and the key point representing the left eye P0 at the next moment when the pedestrian's head lands ’ and P1 ’ The absolute value of the difference between the ordinates |y1 ’ -y0 ’ |y1-y0| is smaller than the absolute value of the difference between the ordinates of P0 and P1 at the previous moment. If the above conditions are met, it is considered that the pedestrian's head has landed headfirst. (2) Landing on the right side of the head: Extract the coordinate information of the five key points of the pedestrian’s head during the landing process. If the absolute value of the difference between the ordinates of P0 and P1 at the current moment |y1-y0| is greater than P0 at the next moment during the landing process, ’ and P1 ’ The absolute value of the difference between the ordinates |y1 ’ -y0 ’ | Large case, while P O The vector composed of P1 and P0 ’ and P1 ’ When the angle between the composed vectors gradually decreases and the final result is within the range of 0°-45° after calculation, it is determined that the right side of the head has landed. (3) Landing on the left side of the head: Extract the coordinate information of the five key points of the pedestrian's head during landing. If the absolute value of the difference between the ordinates of P0 and P1 at the current moment |y1-y0| is equal to P0 at the next moment, ’ and P1 ’ The absolute value of the difference between the ordinates |y1 ’ -y0 ’ |, and at the same time by P O When the angle between the vector formed by P0 and P1 and the vector formed by P0 and P2 is 0°, and the horizontal coordinates of P0, P1, and P2 are consistent, and the vertical coordinate of P0 is greater than the vertical coordinates of P1 and P2, it is determined that the left side of the head has landed. (4) The back of the head touches the ground: The coordinate information of the five key points of the pedestrian’s head during landing is extracted. If the absolute value of the difference between the ordinates of P0 and P1 at the current moment |y1-y0| is equal to P0 at the next moment, ’ and P1 ’ The absolute value of the difference between the ordinates |y1 ’ -y0 ’ |, and at the same time, the angle between the vector composed of P0 and P1 and the vector composed of P0 and P2 is 0°, and at the same time, the vertical coordinates of the coordinate points P0, P1 and P3 are equal and P0 and P1 coincide, it is determined that the back of the head has landed.

6. The method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle according to claim 1, characterized in that: Step 3 The process of extracting the key points of the pedestrian's posture from the key frame is as follows: Step 3.1: Based on the obtained candidate frame, use the single pose estimator SPPE in the Alphapose algorithm to perform in-depth analysis on the human body area in the candidate frame to obtain the key human body area; Step 3.2: Use the spatial transformer network STN to analyze key human body areas; Step 3.3: Use convolutional neural network to extract features of key human body areas; Step 3.4: Use SPPE to process the extracted features to obtain the posture key points in the key human body area, use the posture guided region generator PGPG to process the posture key points, and correct the posture key points with large deviations; Step 3.5: Use the reverse spatial variation network SDTN to map the pose key points to key frames according to body parts.

7. The method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle according to claim 1, characterized in that: The process of extracting the human body area and candidate frame in the image in step 2 is: Step 2.1: Take the key frame obtained in step 1 as the input image, transform the input key frame image, and locate the pedestrian in the image at the center of the detection frame; Step 2.2: Use the region proposal network RPN to identify the human body area in all key frames, and the region proposal network (RPN) outputs the candidate box; Step 2.3: Use non-maximum suppression to perform confidence detection on the candidate boxes to eliminate redundant candidate boxes.

8. The method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle according to claim 7, characterized in that: The input key frame image is transformed using the spatial transformer network STN.

9. The method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle according to claim 7, characterized in that: In step 2.3, non-maximum suppression (NMS) is used to perform confidence detection on the candidate box, and non-maximum suppression includes confidence elimination and distance elimination.

10. The method for determining the direction of a pedestrian's head when landing after a collision between a pedestrian and a vehicle according to claim 1, characterized in that: Step 1 uses the YoloV5-CBAM algorithm to extract key frames containing pedestrian information from the acquired vehicle external image data.

Images