A Map-Semantic Guided Monocular Pedestrian Indoor Localization Prediction Method

Through the map semantic guidance of monocular pedestrian indoor positioning prediction method, combined with machine learning network and dynamic data grid modeling, the pedestrian occlusion relationship is solved, and the problem of pedestrian trajectory interruption and missing positioning results in pedestrian positioning in indoor monocular camera is solved, and the problem of pedestrian trajectory interruption and positioning results is missing in pedestrian positioning of indoor monocular cameras is realized. Continuity and integrity of pedestrian trajectory.

CN114937060BActive Publication Date: 2024-11-12CHENGDU BEIDOU KANGYAO INNOVATION TECHNOLOGY CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202210449486.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-26
Publication Date
2024-11-12
Estimated Expiration
2042-04-26

AI Technical Summary

Technical Problem

In the indoor monocular camera pedestrian positioning, pedestrians suddenly change or leave the video surveillance area, resulting in the missing positioning results and interruption of trajectory. The existing prediction model is not effective in complex scenarios, making it difficult to ensure continuous and smooth positioning effect.

Method used

Using a map semantic-guided monocular pedestrian indoor positioning prediction method, the loosely coupled machine learning network integrates the skeleton detection model, visual behavior positioning model and video pedestrian tracking model, dynamic data raster modeling, calculate pedestrian candidate position sequence and score, The highest-rated sequence is selected as the prediction result to process the occlusion relationship between pedestrians.

Benefits of technology

It effectively solves the problems of pedestrian trajectory interruption and missing positioning results, ensures the integrity and continuity of pedestrian trajectory in the monitoring area, and has improved effects especially when long-term positioning is missing in indoor monitoring space.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114937060B_ABST
    Figure CN114937060B_ABST
Patent Text Reader

Abstract

The invention discloses a monocular pedestrian indoor positioning prediction method guided by map semantics, which relates to the field of indoor positioning technology. The method integrates machine learning models such as a skeleton detection model, a pedestrian positioning model and a pedestrian tracking model in a loosely coupled manner to obtain the position information of pedestrians in a monitoring space. Based on the position information when the pedestrian positioning is lost, in combination with other pedestrian positions and map semantic information, a candidate position sequence prediction algorithm guided by an entrance and exit is used to obtain a candidate position sequence set on the basis of a Kalman filter model integrating social forces. Finally, the occlusion relationship between pedestrians is modeled, and then each candidate position sequence is scored, and the one with the highest score is selected as the final prediction result, thereby solving the problems of missing positioning results and interruption of pedestrian trajectories due to occlusion, missed detection and the like, and effectively ensuring the integrity and continuity of the pedestrian trajectories in the monitoring area, and having a good effect on the situation of long-term positioning loss in the indoor monitoring space.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of indoor positioning technology, and in particular to a monocular pedestrian indoor positioning prediction method guided by map semantics. Background Art

[0002] As a type of indoor positioning, indoor monocular camera pedestrian positioning has the characteristic of suddenness. The suddenness means that pedestrians may suddenly appear or leave the video surveillance area in different motion states, and sudden changes in the environment and pedestrians may lead to the loss of positioning results. How to ensure continuous and smooth positioning effects in monocular vision positioning remains to be studied.

[0003] Existing studies have mostly used historical positions to predict future positions, which have good prediction effects in large-scale simple scenarios. However, recent studies have shown that the interactions and motion states between pedestrians are too complex and may change their behaviors in an instant. In most cases in complex scenarios, various prediction models are not significantly better than the prediction of the uniform motion model. In indoor monitoring spaces, pedestrian movements change more frequently, so predictions based on historical positions are not applicable.

[0004] The present invention can solve the problems of missing positioning results and interrupted pedestrian trajectories due to factors such as occlusion and missed detection, effectively ensure the integrity and continuity of pedestrian trajectories in the monitoring area, and has a better improvement effect on the situation of long-term positioning loss in indoor monitoring spaces. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a monocular pedestrian indoor positioning prediction method guided by map semantics, which is characterized by comprising the following steps:

[0006] S1. Integrate the machine learning network in a loosely coupled manner. The machine learning network includes a skeleton detection model, a visual behavior positioning model MonoLoco, and a video pedestrian tracking model FairMOT. The positions and identity numbers of pedestrians in the monitoring space are associated and tracked to obtain the position information of pedestrians, including the position information when the pedestrian positioning is lost.

[0007] S2, dynamic data grid modeling of pedestrians and obstacles in the monitoring space, and construction of a grid map in the monitoring space;

[0008] S3, combining the position information of the pedestrian when the positioning is lost obtained in step S1 and the grid map semantic information obtained in step S2, calculating a candidate position sequence set of the pedestrian;

[0009] S4. Model the occlusion relationship between pedestrians. Based on the occlusion judgment of pedestrians, score all candidate position sequences in the candidate position sequence set, sort them according to the scores, and select the candidate position sequence with the highest score as the final position prediction result.

[0010] The technical solution further defined in the present invention is:

[0011] Furthermore, the loose coupling method of the visual behavior positioning model MonoLoco and the video pedestrian tracking model FairMOT in step S1 is set as follows: the pedestrian positioning result is obtained in the video frame through the visual behavior positioning model MonoLoco, and the pedestrian tracking result is obtained through the video pedestrian tracking model FairMOT, and the overlap of the pedestrian positioning result and the pedestrian bounding box (bbox) in the pedestrian tracking result is calculated. When the overlap between the two is lower than a pre-set threshold, it is regarded as pedestrian positioning missing.

[0012] In the aforementioned monocular pedestrian indoor positioning prediction method guided by map semantics, the calculation function of the overlap between the pedestrian positioning result and the pedestrian bounding box (bbox) in the pedestrian tracking result is set to the DIoU loss function;

[0013] The calculation method of the overlap between the above two is set as follows: the pedestrian tracking result is used as the basic result, and the pedestrian positioning result is used as the comparison result to determine whether the pedestrian tracking result can be matched with the pedestrian positioning result. If no match is found, it is considered that the pedestrian positioning is missing.

[0014] In the above-mentioned monocular pedestrian indoor positioning prediction method guided by map semantics, the method for obtaining pedestrian position information in step S1 includes the following sub-steps:

[0015] S1.1. Arrange a camera or connect an existing camera device to access the video stream and input the predicted image;

[0016] S1.2. Associate the pedestrian prediction boxes (bboxes) in the prediction results of the visual behavior positioning model MonoLoco and the video pedestrian tracking model FairMOT, and merge the pedestrian tracking results with the positioning results;

[0017] S1.3. Deploy a loosely coupled machine learning network to predict the incoming video and record the location information of pedestrians.

[0018] In the above-mentioned monocular pedestrian indoor positioning prediction method guided by map semantics, the algorithm of the candidate position sequence set of pedestrian prediction in step S3 is set to a candidate position sequence prediction algorithm guided by entrances and exits, and the algorithm includes the following sub-steps:

[0019] S3.1. Taking the position of the pedestrian when the positioning is lost as the starting point and the center position of the entrance and exit of the monitoring space as the end point, the path planning method is used to extract the initial candidate path of the pedestrian;

[0020] S3.2. Correct and optimize the initial candidate path to obtain a candidate position sequence, traverse each entrance and exit of the monitoring space in turn, and calculate the candidate position sequence set.

[0021] In the above-mentioned monocular pedestrian indoor positioning prediction method guided by map semantics, the path planning method in step S3.1 is set to the A-Star algorithm.

[0022] In the above-mentioned monocular pedestrian indoor positioning prediction method guided by map semantics, the correction and optimization method of the initial candidate path in step S3.2 includes the following sub-steps:

[0023] S3.2.1. Construct a social force model for pedestrians. The social force model includes four social force components and one social force resultant. The four social force components are respectively set as motion behavior force, target attraction force, obstacle repulsion force and pedestrian interaction force;

[0024] S3.2.2, determine a fixed spacing according to the grid size of the grid map, select a local path on the initial candidate path of the pedestrian described in step 3.1 with this fixed spacing, and use this local path as the local target of the Kalman filter model integrating social forces to predict the next local target;

[0025] S3.2.3. Repeat the correction until the end of the pedestrian's initial candidate path is reached to obtain a smooth candidate position sequence, and then traverse each entrance and exit of the monitoring space in turn to calculate the candidate position sequence set.

[0026] In the above-mentioned monocular pedestrian indoor positioning prediction method guided by map semantics, the modeling method of the occlusion relationship between pedestrians in step S4 is set as:

[0027] At time t, select position p in the candidate position sequence i and the actual pedestrian position p j , the camera position is o, when the vector (pi-o) and the vector (p j -o) and the cosine of the angle between pi and p j When the Euclidean distance between the pedestrians meets the threshold condition at the same time, the pedestrian is judged to be blocked.

[0028] In the above-mentioned monocular pedestrian indoor positioning prediction method guided by map semantics, the modeling method of the occlusion relationship between pedestrians in step S4 is set as:

[0029] The camera's ray model is used to evaluate whether the pedestrian is blocked. The occlusion judgment expression of the model is described as follows:

[0030] At time t, position p in the candidate position sequence i and the actual pedestrian position p j , the camera position is o, and the occlusion relationship is expressed as:

[0031]

[0032] where dist(·) is the position p i and position p j Euclidean distance function, angle(·) is a vector (p i -o) and vector (p j -o), a and b are the thresholds of distance and angle respectively. Only when both threshold conditions are met at the same time, the pedestrian is judged to be blocked, where threshold a=1 and threshold b=2.

[0033] In the above-mentioned monocular pedestrian indoor positioning prediction method guided by map semantics, the scoring method of the candidate position sequence in step S4 is set as:

[0034] According to the occlusion judgment of pedestrians, the number of occluded pedestrians, the number of missed positions and the total number of positions in the candidate position sequence are obtained. The number of occluded pedestrians is m, the number of missed positions is n, and the total number of positions is s=m+n. The candidate position sequence score is pre The calculation formula is

[0035]

[0036] Where ω is the weight adjustment factor, take ω = 1; score pre It is proportional to m and inversely proportional to n. When all positions are blocked, the maximum value e is obtained.

[0037] The beneficial effects of the present invention are:

[0038] (1) In the present invention, a single-camera pedestrian continuous localization algorithm assisted by map semantic information is proposed, which integrates deep learning models of various tasks in a loosely coupled manner and combines the pedestrian's identity number and location information; then, based on the grid map of the monitoring space, a candidate position sequence prediction algorithm guided by the entrance and exit is used to calculate the candidate position sequence set of the pedestrian, and the sequence set is scored according to the occlusion relationship between pedestrians, and the sequence set with the highest score is selected as the prediction result, finally realizing the prediction of pedestrian trajectory in indoor situations such as pedestrian occlusion and missed detection;

[0039] (2) In the present invention, the visual behavior localization model MonoLoco is a visual model specifically for monocular pedestrian localization. The model infers the depth of pedestrians by obtaining skeleton information from RGB images, and then obtains the geographical location of pedestrians. In addition, it also estimates the uncertainty of the localization result, that is, the model feeds back the accuracy evaluation index while obtaining the localization result;

[0040] (3) In the present invention, the video pedestrian tracking model FairMOT first feeds the input image into an encoder-decoder network to extract a high-resolution feature map, then adds two simple parallel task heads to predict the pedestrian bounding box (bbox) and ReID features respectively, and finally extracts the features at the center of the predicted endpoint for bounding box temporal connection. FairMOT has shown good results on multiple data sets, and related studies have also verified that the model has good generalization in other application scenarios;

[0041] (4) In the present invention, the DIoU loss function is based on the IoU loss function (intersection over union ratio), taking into account the distance, overlap rate and scale, so that the calculation of the pedestrian bounding box (bbox) overlap is more accurate;

[0042] (5) In the present invention, the path planning method is set to the A-Star algorithm, so that the entire path planning method has the advantages of being fast, simple and highly interpretable. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0044] Figure 2 Schematic diagram of the principle of the DIoU loss function in the present invention;

[0045] Figure 3 is a schematic diagram of a candidate position sequence set in the present invention;

[0046] Figure 4 It is a schematic diagram of the ray model of the monitoring space in the present invention. DETAILED DESCRIPTION

[0047] This embodiment provides a monocular pedestrian indoor positioning prediction method guided by map semantics, such as Figure 1 As shown, the following steps are included:

[0048] S1. Integrate the machine learning network in a loosely coupled manner. The machine learning network includes a skeleton detection model, a visual behavior positioning model MonoLoco, and a video pedestrian tracking model FairMOT. Associate and track the positions and identity numbers of pedestrians in the monitoring space. The identity number is set as follows: The machine learning network automatically numbers the pedestrians in the monitoring space in sequence and obtains the position information of the pedestrians, including the position information when the pedestrian positioning is lost;

[0049] In the video frame, the pedestrian positioning result is obtained through the visual behavior positioning model MonoLoco, and the pedestrian tracking result is obtained through the video pedestrian tracking model FairMOT. The overlap of the pedestrian bounding box (bbox) in the pedestrian positioning result and the pedestrian tracking result is calculated. The calculation function is set to the DIoU loss function, which can take into account the distance, overlap rate and scale. The matching principle is as follows: Figure 2 As shown; when the overlap between the two is lower than the pre-set threshold, it is considered that the pedestrian positioning is missing. The specific value of the threshold can be adjusted according to the situation. When the scene is densely populated with pedestrians, increasing the threshold can achieve the effect of reducing invalid matches. In this embodiment, the threshold is set to 0.5;

[0050] From the elements in the pedestrian positioning result and the pedestrian tracking result, it can be found that both contain the pedestrian bounding box (bbox). The pedestrian bounding box (bbox) of the same pedestrian should be highly overlapped. Therefore, the tracking result and the positioning result can be merged by associating the pedestrian bounding boxes (bbox) of the two.

[0051] Considering that the video pedestrian tracking model FairMOT has better pedestrian detection effect than the visual behavior positioning model MonoLoco, the pedestrian tracking result is used as the basic result and the pedestrian positioning result is used as the comparison result to determine whether the pedestrian tracking result can match the pedestrian positioning result. If it cannot match, it is considered that the pedestrian positioning is missing;

[0052] The visual behavior positioning model MonoLoco is a visual model specifically designed for monocular pedestrian positioning. The model infers the depth of pedestrians by obtaining skeleton information from RGB images, and then obtains the geographical location of pedestrians. In addition, it also estimates the uncertainty of the positioning results, that is, the model feeds back the accuracy evaluation index while obtaining the positioning results;

[0053] The video pedestrian tracking model FairMOT first feeds the input image into an encoder-decoder network to extract high-resolution feature maps, then adds two simple parallel task heads to predict pedestrian bounding boxes (bboxes) and ReID features respectively, and finally extracts the features at the center of the predicted endpoint for bounding box temporal connection. FairMOT has shown good results on multiple datasets, and related research has also verified that the model has good generalization in other application scenarios;

[0054] The method for obtaining pedestrian location information includes the following sub-steps:

[0055] S1.1. Arrange a camera or connect an existing camera device to access the video stream and input the predicted image;

[0056] S1.2. Associate the pedestrian prediction boxes (bboxes) in the prediction results of the visual behavior positioning model MonoLoco and the video pedestrian tracking model FairMOT, and merge the pedestrian tracking results with the positioning results;

[0057] S1.3. Deploy a loosely coupled machine learning network to predict the incoming video and record the location information of pedestrians.

[0058] S2. Dynamic data grid modeling is performed on pedestrians and obstacles in the monitoring space to construct a grid map in the monitoring space.

[0059] S3, combining the position information of the pedestrian when the positioning is lost obtained in step S1 and the grid map semantic information obtained in step S2, using the candidate position sequence prediction algorithm guided by the entrance and exit to calculate the candidate position sequence set of the pedestrian;

[0060] The entrance-exit guided candidate position sequence prediction algorithm includes the following sub-steps:

[0061] S3.1. Taking the position where the pedestrian positioning is lost as the starting point and the center position of the entrance and exit of the monitoring space as the target, the path planning method is used to extract the initial candidate path of the pedestrian. The path planning method is set to the A-Star algorithm;

[0062] S3.2, modify and optimize the initial candidate path to obtain a candidate position sequence, traverse each entrance and exit of the monitoring space in turn, and calculate the candidate position sequence set;

[0063] The correction and optimization method of the initial candidate path in step S3.2 includes the following sub-steps:

[0064] S3.2.1. Construct a social force model for pedestrians. The social force model includes four social force components and one social force resultant. The four social force components are respectively set as motion behavior force, target attraction force, obstacle repulsion force and pedestrian interaction force;

[0065] S3.2.2, determine a fixed spacing according to the grid size of the grid map, select a local path on the initial candidate path of the pedestrian described in step 3.1 with this fixed spacing, and use this local path as the local target of the Kalman filter model integrating social forces to predict the next local target;

[0066] S3.2.3, loop correction until the end of the pedestrian's initial candidate path, obtain a smooth candidate position sequence, and then traverse each entrance and exit of the monitoring space in turn to calculate the candidate position sequence set;

[0067] like Figure 3 As shown, the solid lines represent the trajectory segments composed of the real positions, such as A0 and B, and the dotted lines represent the trajectory sets formed by the candidate positions predicted by the algorithm, such as A1, A2 and A3, where A1 corresponds to entrance 1, A2 corresponds to entrance 3, and A3 corresponds to entrance 4.

[0068] S4, modeling the occlusion relationship between pedestrians, scoring all candidate position sequences in the candidate position sequence set based on the occlusion judgment of pedestrians, and sorting them according to the scores, and selecting the candidate position sequence with the highest score as the final position prediction result;

[0069] The modeling method of the occlusion relationship between pedestrians in step S4 is set as:

[0070] At time t, select position p in the candidate position sequence i and the actual pedestrian position p j , the camera position is o, when the vector (p i -o) and vector (p j -o) and the cosine of the angle between i With p j When the Euclidean distance between the pedestrians meets the threshold condition at the same time, the pedestrian is judged to be blocked;

[0071] Considering the small range of indoor monitoring space and the close distance between pedestrians and cameras, visual detection of indoor pedestrians is usually not easy to miss. However, pedestrians close to each other are prone to occlusion, resulting in the loss of pedestrian positions. Whether pedestrians are occluded is evaluated based on the camera's ray model. The ray model also considers the distance between pedestrian positions and the angle between the two positions and the camera. The occlusion judgment expression of the model is described as follows:

[0072] At time t, position p in the candidate position sequence i and the actual pedestrian position p j , the camera position is o, and the occlusion relationship is expressed as:

[0073]

[0074] where dist(·) is the position p i and position p j Euclidean distance function, angle(·) is a vector (p i -o) and vector (p j -o), a and b are the thresholds of distance and angle respectively. Only when both threshold conditions are met at the same time, the pedestrian is judged to be blocked, where threshold a = 1 and threshold b = 2;

[0075] Figure 4 The ray model of the monitoring space is shown, where A, B, C, D, and E are the positions of pedestrians at the same time, and B is blocked by C and is represented by a dotted line;

[0076] The scoring method for the candidate position sequence in step S4 is set as:

[0077] According to the occlusion judgment of pedestrians, the number of occluded positions, the number of missed positions and the total number of positions of pedestrians in the candidate position sequence are obtained. The final scoring basis of the position sequence is: assuming that the loss of pedestrian positions is caused by missed detection or occlusion, the more occluded positions in the candidate position sequence and the fewer missed positions, the higher the score of the candidate position sequence;

[0078] Assume that in the candidate position sequence, the number of blocked positions is m, the number of missed positions is n, the total number of positions is s = m + n, and the candidate position sequence score is pre The calculation formula is

[0079]

[0080] Where ω is the weight adjustment factor, which changes according to the monitoring scenario, and takes ω = 1; score pre It is proportional to m and inversely proportional to n. When all positions are blocked, the maximum value e is obtained.

[0081] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope required by the present invention.

Claims

1. A monocular pedestrian indoor positioning prediction method guided by map semantics, characterized by: The following steps are included S1. Integrate the machine learning network in a loosely coupled manner. The machine learning network includes a skeleton detection model, a visual behavior positioning model MonoLoco, and a video pedestrian tracking model FairMOT. The positions and identity numbers of pedestrians in the monitoring space are associated and tracked to obtain the position information of pedestrians, including the position information when the pedestrian positioning is lost. S2, dynamic data grid modeling of pedestrians and obstacles in the monitoring space, and construction of a grid map in the monitoring space; S3, combining the position information of the pedestrian when the positioning is lost obtained in step S1 and the grid map semantic information obtained in step S2, calculating a candidate position sequence set of the pedestrian; S4, modeling the occlusion relationship between pedestrians, scoring all candidate position sequences in the candidate position sequence set based on the occlusion judgment of pedestrians, and sorting them according to the scores, and selecting the candidate position sequence with the highest score as the final position prediction result; The loose coupling method of the visual behavior positioning model MonoLoco and the video pedestrian tracking model FairMOT in step S1 is set as follows: the pedestrian positioning result is obtained in the video frame by the visual behavior positioning model MonoLoco, and the pedestrian tracking result is obtained by the video pedestrian tracking model FairMOT, and the overlap between the pedestrian positioning result and the pedestrian bounding box (bbox) in the pedestrian tracking result is calculated. When the overlap between the two is lower than a pre-set threshold, it is considered that the pedestrian positioning is missing; The calculation function of the overlap between the pedestrian localization result and the pedestrian bounding box (bbox) in the pedestrian tracking result is set to the DIoU loss function; The calculation method of the overlap between the above two is set as follows: the pedestrian tracking result is used as the basic result, and the pedestrian positioning result is used as the comparison result. It is judged whether the pedestrian tracking result can be matched with the pedestrian positioning result. If it cannot be matched, it is regarded as the pedestrian positioning is missing. The algorithm for the candidate position sequence set of pedestrian prediction in step S3 is set to the candidate position sequence prediction algorithm guided by the entrance and exit, and the algorithm includes the following sub-steps: S3.

1. Taking the position of the pedestrian when the positioning is lost as the starting point and the center position of the entrance and exit of the monitoring space as the end point, the path planning method is used to extract the initial candidate path of the pedestrian; S3.2, modify and optimize the initial candidate path to obtain a candidate position sequence, traverse each entrance and exit of the monitoring space in turn, and calculate the candidate position sequence set; The correction optimization method of the initial candidate path in step S3.2 includes the following sub-steps: S3.2.

1. Construct a social force model for pedestrians. The social force model includes four social force components and one social force resultant. The four social force components are respectively set as motion behavior force, target attraction force, obstacle repulsion force and pedestrian interaction force; S3.2.2, determine a fixed spacing according to the grid size of the grid map, select a local path on the initial candidate path of the pedestrian described in step 3.1 with this fixed spacing, and use this local path as the local target of the Kalman filter model integrating social forces to predict the next local target; S3.2.

3. Repeat the correction until the end of the pedestrian's initial candidate path is reached to obtain a smooth candidate position sequence, and then traverse each entrance and exit of the monitoring space in turn to calculate the candidate position sequence set.

2. The map semantics-guided monocular pedestrian indoor positioning prediction method according to claim 1, characterized in that: The method for obtaining pedestrian location information in step S1 includes the following sub-steps S1.1, arranging a camera or connecting an existing camera device to access a video stream, and inputting a predicted image; S1.

2. Associate the pedestrian prediction boxes (bboxes) in the prediction results of the visual behavior positioning model MonoLoco and the video pedestrian tracking model FairMOT, and merge the pedestrian tracking results with the positioning results; S1.

3. Deploy a loosely coupled machine learning network to predict the incoming video and record the location information of pedestrians.

3. The map semantics-guided monocular pedestrian indoor positioning prediction method according to claim 1, characterized in that: The path planning method in step S3.1 is set to the A-Star algorithm.

4. The map semantics-guided monocular pedestrian indoor positioning prediction method according to claim 1, characterized in that: The modeling method of the occlusion relationship between pedestrians in step S4 is set as follows: At time t, select position p in the candidate position sequence i and the actual pedestrian position p j , the camera position is o, when the vector (p i -o) and vector (p j -o) and the cosine of the angle between i With p j When the Euclidean distance between the pedestrians meets the threshold condition at the same time, the pedestrian is judged to be blocked.

5. The map semantics-guided monocular pedestrian indoor positioning prediction method according to claim 4, characterized in that: The modeling method of the occlusion relationship between pedestrians in step S4 is set as follows: The camera's ray model is used to evaluate whether the pedestrian is blocked. The occlusion judgment expression of the model is described as follows: At time t, position p in the candidate position sequence i and the actual pedestrian position p j , the camera position is o, and the occlusion relationship is expressed as where dist(·) is the position p i and position p j Euclidean distance function, angle(·) is a vector (p i -o) and vector (p j -o), a and b are the thresholds of distance and angle respectively. Only when both threshold conditions are met at the same time, the pedestrian is judged to be blocked, where threshold a=1 and threshold b=2.

6. The map semantics-guided monocular pedestrian indoor positioning prediction method according to claim 4, characterized in that: The scoring method of the candidate position sequence in step S4 is set as: According to the occlusion judgment of pedestrians, the number of occluded pedestrians, the number of missed positions and the total number of positions in the candidate position sequence are obtained. The number of occluded pedestrians is m, the number of missed positions is n, and the total number of positions is s=m+n. The candidate position sequence score is pre The calculation formula is Where ω is the weight adjustment factor, take ω = 1; score pre It is proportional to m and inversely proportional to n. When all positions are blocked, the maximum value e is obtained.

Citation Information

Patent Citations

  • Joint reasoning-based video multi-target tracking method

    CN103699908A

  • Target tracking method and system based on video monitoring

    CN106558069A

  • Indoor pedestrian positioning method based on monocular vision

    CN112258571A

  • Pedestrian Tracking Method and Pedestrian Tracking Device

    US20100002908A1