An intelligent indoor navigation method based on image target detection and tracking

Through image object detection and tracking technology and multi-camera work in collaboration, combined with the Astar algorithm, the problems of low indoor navigation accuracy and low fault tolerance in the existing technology are solved, real-time navigation and precise path planning in complex environments are realized.

CN114674323BActive Publication Date: 2024-11-15BEIJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210382299.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-12
Publication Date
2024-11-15
Estimated Expiration
2042-04-12

AI Technical Summary

Technical Problem

Existing indoor navigation technology is difficult to achieve real-time tracking and accurate navigation in multi-target environments, especially in crowded and complex places, resulting in low navigation accuracy and reduced fault tolerance.

Method used

An intelligent indoor navigation method based on image object detection and tracking is adopted, and a monocular camera and QR code are used to achieve user identity matching, and the navigation coverage is expanded through the joint multi-camera, and the shortest path is planned in combination with the Astar algorithm.

Benefits of technology

Real-time user tracking and accurate navigation in a multi-target environment are achieved, navigation errors are reduced, navigation accuracy and fault tolerance are improved, and the equipment requirements are simple, low cost is suitable for a wide range of application scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114674323B_ABST
    Figure CN114674323B_ABST
Patent Text Reader

Abstract

The present invention discloses an intelligent indoor navigation method based on image target detection and tracking, which belongs to the field of camera positioning and tracking. First, for a number of users queuing to enter the room, when each user passes each camera, the host computer automatically matches the ID and tracking detection frame under each camera for each user; and locates the actual physical space coordinates of each user in each camera; then, the distance is calculated for each user in the queue, and the user ID with the smallest distance from its own camera is selected, and the user QR code is generated by matching; until each user generates their own QR code; the starting and ending coordinates of each user are transmitted to the host computer respectively, the shortest path is planned and returned to each user end; finally, when each user moves along the shortest path from the starting point and enters the cross-camera area, the cross-camera pedestrian tracking technology is used to transmit and track the user data and update the navigation results in real time; the present invention effectively expands the coverage area and is suitable for monitoring places.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of camera positioning and tracking, pedestrian recognition and path planning, and specifically is an intelligent indoor navigation method based on image target detection and tracking. Background Art

[0002] GPS technology is an outdoor positioning and navigation system that was first used by the public. It is relatively mature and has become an essential tool for outdoor positioning and navigation services. However, satellite positioning signals are weak and easily affected by environmental factors such as obstruction, so they cannot play their technical advantages in indoor positioning.

[0003] With the development of positioning and navigation technology in today's era, scientific research teams have realized combined navigation services based on ubiquitous wireless networks and multi-format big data integration by using mobile communication network platforms; while existing indoor positioning technologies such as WLAN, WSN and other systems, although they have achieved efficient regional indoor positioning, require the deployment of a large number of nodes, have high signal coverage costs, and are not conducive to wide-area positioning applications.

[0004] Therefore, the development goal is to improve the existing technology and combine and utilize it to simplify the indoor navigation problem. There are various structures in the indoor scene, with dense personnel and a wide variety of targets (objects). If the real-time tracking of users cannot be completed, the process cannot be controlled, which will lead to a decrease in fault tolerance; if the instructions during navigation are simple, the navigation accuracy will also be greatly reduced. The traditional indoor navigation method has limited capabilities and cannot take into account many aspects; for example, document 1: application number 201621245486.6 proposes an indoor vehicle positioning and navigation system based on video image processing, and the moving vehicle information obtained by the image acquisition device is sent to the server through the communication device. The background server sends the location of the user vehicle to the vehicle user terminal after image processing. For example, document 2: application number 201310205587.5 proposes an indoor navigation system and its indoor navigation method, which realizes indoor positioning based on QR code, solves the problem of inaccurate indoor positioning in the existing technology, and improves the accuracy of indoor positioning; navigation is carried out according to the guidance of the color sequence to improve the rationality and popularity of the indoor navigation system. For example, document 3: application number 200610063715.7 proposes a map recording method and device for tracking navigation, which saves the points of interest closest to the current location of the controlled terminal in an electronic map database, marks the points of interest and highlights the points of interest, so that the main control terminal can clearly and conveniently understand the current location of the controlled terminal.

[0005] However, the shortcoming of document (1) is that it does not explain how to match a specific vehicle user terminal with the corresponding vehicle in the captured image when the image acquisition device captures multiple vehicles at the same time. Therefore, when the image acquisition device captures images of multiple vehicles and multiple vehicle user terminals send requests to the server at the same time, matching errors may occur, resulting in the user's information such as location information, selected destination coordinates, and returned navigation results being mistransmitted.

[0006] The shortcoming of document (2) is that it is limited in applicable scenarios. The QR code identification information of the QR code tag is bound to the target location information, so the system can only locate the user at the time of scanning the QR code, and cannot complete real-time tracking; and during the user's navigation process, the user can only reach the destination through ground route guidance, and the navigation accuracy is low. When the user is in a place with complex routes and dense crowds such as a library or a shopping mall, it is difficult to complete effective path finding through the navigation system. At the same time, when there is a problem with the user's forward path, it is difficult for the system to make prompts and adjustments in time, and the fault tolerance rate is low.

[0007] The shortcoming of document (3) is that if indoor tracking and navigation uses GPS positioning, it is easily affected by location, weather or other factors in most scenarios, and the effect is not ideal; the civilian GPS accuracy can reach 10 meters, but it cannot play its advantages in complex and diverse indoor places. In addition, most indoor scenes have a large flow of people and a small space, so accurately tracking users is also a problem.

[0008] Nowadays, the accuracy of wireless positioning has been improved, but factors such as obstruction by indoor obstacles may cause inaccurate user positioning, resulting in large errors in navigation results. Summary of the invention

[0009] In view of the above problems, the present invention provides an intelligent indoor navigation method based on image target detection and tracking, which uses QR codes to achieve user identity matching, uses monocular camera positioning technology to determine the user's physical space coordinates, and uses multiple cameras to expand the navigation coverage range, completing navigation from the user coordinates to the destination coordinates while continuously tracking the user. Compared with the existing solutions, the equipment requirements of the present invention are simple, and only cameras and servers can meet the requirements. Therefore, the application cost is low, the applicability is strong, and the application scenarios are wide.

[0010] The intelligent indoor navigation method for image target detection and tracking specifically includes:

[0011] Step 1: For several users queuing to enter the room, when each user passes by each monocular camera in the room, the host computer automatically assigns each user an ID corresponding to the camera and a tracking detection frame;

[0012] The id is a random number. When user a passes by several cameras in the room, an id and a tracking detection frame bound to the id will be assigned to the user. Similarly, each user will be reassigned an id and a corresponding tracking detection frame when passing by each camera.

[0013] Step 2: Using positioning technology and combining the tracking detection frame of each user, determine the actual physical space coordinates of each user within the shooting range of each camera;

[0014] The specific steps are:

[0015] Step 201: measure and model the actual indoor scene, construct a 2D digital map, and represent it with a two-dimensional matrix;

[0016] In the map, “0” represents an area of ​​one square meter in the actual scene, “1” represents an obstacle, and “2” represents the target location retrieved by the user in the system based on the name;

[0017] Step 202: For the current camera b and user a, calculate the distance D between the user a and the camera b;

[0018] The calculation formula is:

[0019]

[0020] D is the distance from user a to camera b, P is the pixel width of user a in the image taken by camera b, and W is the height of user a.

[0021] Step 203: Use the current camera b to shoot the A4 paper, and use the measurement data to calculate the camera focal length F of the camera b;

[0022]

[0023] P' is the pixel height of A4 paper, D' is the distance from the lens of camera b to the A4 paper, and W' is the height of A4 paper.

[0024] Step 204: Establish a rectangular coordinate system of the actual scene based on the 2D digital map, and convert the pixel coordinates of the four vertices of the tracking detection box of the user a in the picture into the coordinates (x1, y1) of the user a in the actual scene:

[0025] x1=X max -D

[0026]

[0027] X maxis the maximum value of x in the actual scene built by the 2D digital map, bbox[0] is the horizontal coordinate of the upper left corner of the detection box in the coordinate system composed of pixels in the camera image; bbox[2] is the horizontal coordinate of the lower right corner of the detection box in the coordinate system composed of pixels in the camera image;

[0028] Step 205: Similarly, the physical space coordinates of each user under each monocular camera are calculated respectively.

[0029] Step 3: The initial camera in the room calculates the distance between each user in the queue, selects the user with the shortest distance from itself, and uses the user ID to match and generate a user QR code corresponding to the ID; until each user generates a user QR code;

[0030] Step 4: The starting point coordinates and the end point coordinates of each user are transmitted to the host computer respectively, and the shortest path is planned according to the set map, and then returned to each user terminal;

[0031] The shortest path planning is: take the end point coordinates of each user as the target grid, use the Astar algorithm to start moving along the parent node of each grid until returning to the user's starting grid, and then the best path for each user can be obtained.

[0032] Step 5: When each user moves along the shortest path from the starting point and enters the cross-camera area, each user is tracked and the navigation results are updated in real time.

[0033] During the user tracking process, if the user enters the area covered by one camera from the area covered by another camera, the user data is transferred using cross-camera pedestrian tracking technology:

[0034] When user a walks from the shooting range of camera A to the shooting range of camera B, the images captured by the two cameras will partially overlap.

[0035] The user a is randomly assigned an id k in camera A and an id j in camera B, and user a is bound to the id in camera A through a QR code.

[0036] When user a walks into the overlapping area captured by the two cameras, cameras A and B will simultaneously calculate the coordinates of user a in the actual scene. If the coordinates are consistent, id k in camera A and id j in camera B are the same user, and the data bound to the user is transmitted from camera A to camera B. Otherwise, if the coordinates calculated by camera A and camera B are inconsistent, it is considered that id k and id j appearing in the overlapping area are two different people, and the user data will not be transmitted.

[0037] Similarly, when user a continues to walk along the shortest path from camera B and reaches the next camera C, the user coordinates are calculated according to the overlapping area of ​​the two cameras according to the above process, and so on, until the user reaches the target grid and completes the complete data transmission.

[0038] The advantages of the present invention are:

[0039] 1) An intelligent indoor navigation method based on image target detection and tracking uses a monocular camera ranging technology to detect the digital ID of the user closest to the camera, and uses the ID to generate a QR code unique to the user. After the user scans the QR code, the user identity and the digital ID in the system can be matched. When the system provides services to multiple users at the same time, identity matching can ensure that data transmission between different user terminals and the server is not confused.

[0040] 2) An intelligent indoor navigation method based on image target detection and tracking, based on the cross-camera navigation method, realizes multi-camera collaborative work to complete tracking navigation. When a user enters the overlapping shooting area of ​​two cameras, the two cameras calculate the user's physical coordinates respectively. If the physical coordinates are consistent, the two cameras consider it to be the same user and bind the user's data with the id in the new camera. When the user enters the next area, continuous tracking navigation can be completed according to the id in the user's new camera; effectively expand the coverage area and adapt to the monitoring site. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart of an intelligent indoor navigation method based on image target detection and tracking of the present invention;

[0042] Figure 2 A schematic diagram of assigning an ID and a corresponding bounding box to each user detected by the system of the present invention;

[0043] Figure 3 It is the workflow diagram of the yolov+deepsort algorithm of the present invention;

[0044] Figure 4 This is the flow chart of the Astar algorithm of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0046] Existing indoor positioning technologies such as WLAN, WSN, etc., although they achieve efficient regional indoor positioning, require the deployment of a large number of nodes, the signal coverage cost is high, and the coverage area is also limited. The present invention proposes a new idea for indoor navigation implementation. One of the biggest features is that the equipment requirements are simple. Only a camera and a simple server can meet the needs, with the advantages of low cost and fast calculation. It can even consider using the original surveillance cameras in the scene, so it has strong scalability and a wide range of application scenarios; through QR code scanning, the actual user's individual identity and the digital ID in the system are matched, and the cross-camera navigation method is combined to realize the collaborative work of multiple cameras to complete tracking navigation, which reduces the technical difficulty and improves the user experience; it can solve the following three problems about indoor navigation:

[0047] 1) Utilization of existing indoor facilities

[0048] Most of the existing technologies use additional facilities or tools to achieve image acquisition or positioning. However, due to the popularity of indoor monitoring, most application scenarios have set up related equipment. Better use of it can reduce costs and is easy to operate and popularize. In view of this, the present invention uses the existing camera in the scene. When receiving the request for indoor navigation from the front end, the current geographical location of the target and the indoor navigation map corresponding to the current geographical location are obtained through the camera, and the navigation route is calculated according to the target location, so as to realize the efficient use of existing facilities.

[0049] 2) Improvement of user personalized service experience

[0050] In existing indoor navigation technology, when there are multiple options for the same destination, the indoor server will obtain the geographical locations that match the identifiers of each destination, generate an indoor navigation route that includes the current location of the user terminal and each geographical location, and feed all the results back to the indoor map. Although the indoor navigation needs are met, they are complex and not intuitive enough. Based on this, this application simplifies the path planning process, strives to be concise in the processing of multiple identical destinations, selects the optimal solution from multiple navigation routes and feeds it back to the user end; and attempts to integrate user behavior modeling into the preset method. When planning a route, you can choose whether to pass through recommended or frequently browsed and purchased products by users, thereby improving the user experience.

[0051] 3) Balance between cost, efficiency and accuracy

[0052] Since there is no need to set up additional equipment for the system operation, the cost of this application is greatly reduced; the target detection and tracking process effectively uses the camera to determine the location in the indoor map. Compared with Bluetooth or other positioning methods, it can reduce the error of indoor navigation and improve the accuracy; when receiving indoor navigation requests, the navigation route is completed by the preset generation rules of the backend, and finally only the results need to be transmitted back to the front end to provide to the user. The coordination between multiple threads can effectively improve the navigation efficiency. This application combines various factors to achieve a balance of multiple factors.

[0053] Aiming at the problem of inaccurate navigation and positioning in indoor places, the present invention realizes navigation effect through tracking navigation technology based on the existing cameras and two-dimensional digital maps in the places; the overall technical scheme is divided into five processes: the first step is that the user enters the shooting range of the camera, the system detects and tracks the user, determines the user's unique identification and boundary box, and automatically assigns the corresponding ID of each user under the camera; the second step is that the initial camera calculates the distance between each user and the camera, selects the user with the minimum distance, generates the user's QR code using the user ID, and matches the user identity with the assigned ID; the third step is to determine the user's physical space coordinates using the monocular camera positioning technology; the fourth step is to transmit the user's starting point coordinates and end point coordinates to the server side, plan the shortest path according to the map set in the system, and then return to the user side; the fifth step is to track the user and update the navigation results in real time. It is worth noting that in the process of tracking the user, if the user enters the area covered by one camera from the area covered by another camera, the user data is transmitted using the cross-camera pedestrian tracking technology; the camera network is used to cover the indoor space of a large area to achieve the expansion of the indoor navigation coverage area.

[0054] like Figure 1 As shown, the specific steps are as follows:

[0055] Step 1: For several users to be detected who are queuing to enter the room, when each user passes by each monocular camera in the room, the host computer tracks and automatically assigns an ID corresponding to the camera and a tracking detection frame to each user;

[0056] This example uses the existing yolov3+deepsort algorithm, such as Figure 3 As shown, the user is detected and tracked, the user's unique identification and bounding box are determined, and the system assigns an id to each detected user. The id assigned to the user here is a random number, and the user's id is bound to the bounding box of the corresponding user in the system. The specific example is as follows Figure 2 As shown;

[0057] When user a passes through the shooting range of camera a1 in the room, the host computer tracks and automatically assigns a corresponding id1 and a corresponding tracking detection frame to the user a; similarly, when user a passes through the shooting range of camera a2, the host computer tracks and automatically assigns a corresponding id2 and a corresponding tracking detection frame to the user a; similarly, each user passing through each camera will be reassigned an id and a corresponding tracking detection frame.

[0058] Step 2: The initial camera in the room calculates the distance between each user in the queue, selects the user with the shortest distance to itself, and uses the user ID to match and generate a user QR code corresponding to the ID; until each user's own user QR code is generated and displayed on the display screen;

[0059] For example, the initial camera is X, and the queued users are user a, user b, and user c. According to the distance between each user and the camera, user a with the shortest distance is selected. The host computer generates a QR code for user a using the ID assigned to user a by camera X.

[0060] Similarly, after user a leaves, the next user b repeats the above process until each user's own user QR code is generated.

[0061] The user's QR code is used to: first, display the user interface of the indoor navigation application on the mobile phone when the user enters the indoor navigation application; second, match the user individual (or the user's mobile phone) with the ID assigned to it in the system to achieve the matching of physical entity and digital ID;

[0062] Step 3: Using positioning technology and combining the tracking detection frame of each user, determine the actual physical space coordinates of each user within the shooting range of each camera;

[0063] The specific steps are:

[0064] Step 301: measure and model the actual indoor scene, and pre-build a 2D digital map of the actual scene, represented by a two-dimensional matrix;

[0065] As shown in the following figure: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 2 0 1 0 0 0 0 0 0 1 1 1 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0

[0076] Each "0" in the two-dimensional matrix represents an area of ​​one square meter in the actual scene (which can be modified according to the actual situation), "1" represents an obstacle, and "2" represents the location of the target that the user can retrieve by name in the system.

[0077] Step 302: For the current camera b and user a, calculate the distance D between the user a and the camera b;

[0078] The calculation formula is:

[0079]

[0080] D is the distance from user a to camera b, P is the pixel width of user a in the image taken by camera b, and W is the height of user a.

[0081] Step 303: Use the current camera b to shoot the A4 paper, and use the measurement data to calculate the camera focal length F of the camera b;

[0082]

[0083] When measuring, take an A4 paper, move the camera so that the paper just fills the screen, and measure the corresponding data. P' is the pixel height of the A4 paper, D' is the distance from the lens of camera b to the A4 paper, and W' is the height of the A4 paper.

[0084] Step 304: Establish a rectangular coordinate system of the actual scene based on the 2D digital map, and convert the pixel coordinates of the four vertices of the tracking detection box of user a in the picture. Obtain the coordinates (x1, y1) of user a in the actual scene:

[0085] x1=X max -D

[0086]

[0087] X max is the maximum value of x in the actual scene built by the 2D digital map, bbox[0] is the horizontal coordinate of the upper left corner of the detection box in the coordinate system composed of pixels in the camera image; bbox[2] is the horizontal coordinate of the lower right corner of the detection box in the coordinate system composed of pixels in the camera image;

[0088] Step 305: Similarly, the physical space coordinates of each user under each monocular camera are calculated respectively.

[0089] Step 4: The starting point coordinates and the end point coordinates of each user are transmitted to the host computer respectively, and the shortest path is planned according to the set map, and then returned to each user terminal;

[0090] The user's starting point coordinates are obtained from the actual physical space coordinates, and the user's end point coordinates are obtained from the user's input; in scene modeling, the end point coordinates S input by the user have been bound to its actual physical coordinates, and the actual physical coordinates are the user's destination.

[0091] The shortest path planning is: take the end point coordinates of each user as the target grid, use the Astar algorithm to start moving along the parent node of each grid until returning to the user's starting grid, and then the best path for each user can be obtained.

[0092] like Figure 4 As shown, the specific steps are:

[0093] a) Create an open list and initialize the user's starting point coordinates;

[0094] Find the grid with the lowest f(i) in the open list, switch it to the closed list, and call it the current grid.

[0095] The current node value is estimated through the formula:

[0096] f(i)=g(i)+h(i)

[0097] f(i) is the estimated value of the current node, where g(i) is the cost of moving from the starting point to the node, and h(i) is the estimated cost of moving from the current node to the end point.

[0098] b) Determine the eight adjacent grids of the current grid:

[0099] *If it is not passable or is already in the closed list, ignore it;

[0100] * If it is not in the open list, add it; record its f(i), g(i), and h(i) values ​​and use the current cell as the parent node of this cell.

[0101] *If it is already in the open list, compare its value with the original path g(i); if its g(i) value is smaller, change the parent node of this grid to the current grid, recalculate f(i), g(i), and re-sort the open list.

[0102] c) The loop ends when one of the following conditions is met:

[0103] * Add the target grid to the closed list when the target path is found;

[0104] *The target grid is not found and the open list is empty (the target path does not exist at this time).

[0105] Step 5: When each user moves along the shortest path from the starting point and enters the cross-camera area, each user is tracked and the navigation results are updated in real time.

[0106] During the user tracking process, if the user enters the area covered by one camera from the area covered by another camera, the user data is transferred using cross-camera pedestrian tracking technology:

[0107] When user a walks from the shooting range of camera A to the shooting range of camera B, the images captured by the two cameras will partially overlap.

[0108] The user a is randomly assigned an id k in camera A and an id j in camera B, and user a is bound to the id in camera A through a QR code.

[0109] When user a walks into the overlapping area captured by the two cameras, cameras A and B will simultaneously calculate the coordinates of user a in the 2D map of the actual scene. If the coordinates are consistent, id k in camera A and id j in camera B are the same user, and the data bound to the user is transmitted from camera A to camera B. Otherwise, if the coordinates calculated by camera A and camera B are inconsistent, it is considered that id k and id j appearing in the overlapping area are two different people, and the user data will not be transmitted.

[0110] Similarly, when user a continues to walk along the shortest path from camera B and reaches the next camera C, the user coordinates are calculated according to the overlapping area of ​​the two cameras according to the above process, and so on, until the user reaches the target grid and completes the complete data transmission.

[0111] The tracking navigation technology of the present invention is composed of camera positioning, pedestrian tracking and path planning; pedestrian tracking includes: YOLOv3 extraction, DeepSORT deep learning, etc. Indoor places include: shopping malls, underground garages or libraries and other places with poor signals and dense targets. Obtain a specific QR code scan from the indoor place to complete the matching, call the tracking navigation technology for continuous tracking, and use path planning to complete the shortest path indication to complete navigation.

[0112] YOLO is an end-to-end target detection model and a fast and accurate real-time object detection algorithm. Figure 3 As shown, the detection can be roughly divided into two steps: 1. Determine the location of the detection object; 2. Classify the detection object and determine the classification of the detection object. The general process of the work is as follows:

[0113] The backbone network of YOLOv3 is darknet53; the backbone network is mainly divided into three layers, namely input layer, convolution layer and residual network. Each convolution part of darknet53 uses a unique DarknetConv2D structure, L2 regularization is performed during each convolution, and BatchNormalization and LeakyReLU activation functions are performed after convolution. Compared with YOLOv2, YOLOv3 uses a residual network. The residual structure can be directly mapped from the previous feature layer to the next feature layer (jump connection) without convolution, which is helpful for training, feature extraction, and easy optimization.

[0114] DeepSORT introduces the idea of ​​deep learning, mainly using convolutional neural networks to extract features, and uses the Mahalanobis distance and cosine distance between the state predicted by the recursive Kalman filter and the newly arrived state in the tracking process to measure the relationship between the threshold and the two combined indicators to serve the assignment problem, and uses cascade matching to solve the sub-problems in the global allocation. The general process of DeepSORT is as follows:

[0115] First, connect the camera, input the video frame and perform preprocessing: detect and extract deep learning features, calculate the corresponding confidence, and preliminarily screen the candidate frames; then use non-maximum suppression (NMS) to further screen the candidate frames. Next, use Kalman filtering to predict and measure the matching degree, perform cascade matching and IOU (Intersection over Union) regression, and output the results. Finally, repeat the above process for subsequent video frames.

[0116] Among them, DeepSORT uses an 8-dimensional spatial variable to represent the state of the trajectory at a certain moment. These dimensions represent the center coordinates of the candidate box, the ratio of the longitudinal to lateral direction of the candidate box, the height and speed information; and uses the Kalman filter to predict and update the trajectory state, and uses the Mahalanobis distance between the state predicted by the Kalman filter and the newly arrived target state to represent the degree of motion matching; at the same time, in order to reduce the occurrence of logo switching, DeepSORT introduces the apparent features extracted by deep learning to perform matching analysis to supplement the Mahalanobis distance, and uses the minimum cosine distance between the detection and tracking feature vectors as the measure of apparent matching.

[0117] The shortest path planning method is based on the A Star algorithm obtained by optimizing and modifying the Dijkstra algorithm. By repeatedly obtaining the node closest to the starting point and expanding outward, it can have a certain degree of efficiency in the process of path finding and graph traversal.

Claims

1. An intelligent indoor navigation method based on image target detection and tracking, characterized in that: Specifically include: First, for several users queuing to enter the room, when each user passes by each monocular camera in the room, the host computer automatically assigns each user an ID corresponding to the camera and a tracking detection frame; Then, the actual physical space coordinates of each user within the shooting range of each camera are determined by using positioning technology and combining the tracking detection frame of each user; The actual physical space coordinates of each user are calculated as follows: Step 201: measure and model the actual indoor scene, construct a 2D digital map, and represent it with a two-dimensional matrix; Step 202: For the current camera b and user a, calculate the distance D between the user a and the camera b; The calculation formula is: D is the distance from user a to camera b, where P is the pixel width of user a in the image taken by camera b, and W is the height of user a; Step 203: Use the current camera b to shoot the A4 paper, and use the measurement data to calculate the camera focal length F of the camera b; P' is the pixel height of A4 paper, D' is the distance from the lens of camera b to the A4 paper, and W' is the height of A4 paper; Step 204: Establish a rectangular coordinate system of the actual scene based on the 2D digital map, and convert the pixel coordinates of the four vertices of the tracking detection box of the user a in the picture into the coordinates (x1, y1) of the user a in the actual scene: x1=X max -D X max is the maximum value of x in the actual scene built by the 2D digital map, bbox[0] is the horizontal coordinate of the upper left corner of the detection box in the coordinate system composed of pixels in the camera image; bbox[2] is the horizontal coordinate of the lower right corner of the detection box in the coordinate system composed of pixels in the camera image; Step 205: Similarly, calculate the physical space coordinates of each user under each monocular camera respectively; Next, the initial camera in the room is used to calculate the distance of each user in the queue, and the user with the shortest distance to itself is selected. The user ID is used to match and generate a user QR code corresponding to the ID, until a user QR code is generated for each user. Finally, the starting point coordinates and the end point coordinates of each user are transmitted to the host computer respectively, and the shortest path is planned according to the set map, and then returned to each user terminal; when each user moves along the shortest path from the starting point and enters the cross-camera area, each user is tracked and the navigation results are updated in real time; During the user tracking process, if the user enters the area covered by one camera from the area covered by another camera, the user data is transferred using cross-camera pedestrian tracking technology: When user a walks from the shooting range of camera A to the shooting range of camera B, the images captured by the two cameras will partially overlap. The user a is randomly assigned an id k in camera A and an id j in camera B, and the user a is bound to the id in camera A through a QR code; When user a walks into the overlapping area captured by the two cameras, cameras A and B simultaneously calculate the coordinates of user a in the actual scene. If the coordinates are consistent, id k in camera A and id j in camera B are the same user, and the data bound to the user is transmitted from camera A to camera B. Otherwise, if the coordinates calculated by camera A and camera B are inconsistent, it is considered that id k and id j appearing in the overlapping area are two different people, and the user data will not be transmitted. Similarly, when user a continues to walk along the shortest path from camera B and reaches the next camera C, the user coordinates are calculated according to the overlapping area of ​​the two cameras according to the above process, and so on, until the user reaches the target grid and completes the complete data transmission.

2. The intelligent indoor navigation method based on image target detection and tracking as claimed in claim 1, characterized in that: The id is a random number. When user a passes through several cameras in the room, an id and a tracking detection frame bound to the id will be assigned to the user respectively; similarly, each user will be reassigned an id and a corresponding tracking detection frame when passing through each camera.

3. The intelligent indoor navigation method based on image target detection and tracking as claimed in claim 1, characterized in that: The shortest path is planned as follows: the user's starting point coordinates are calculated from the actual physical space coordinates, and the end point coordinates are the actual physical coordinates corresponding to the coordinates input by the user; the end point coordinates of each user are used as the target grid, and the Astar algorithm is used to start moving along the parent node of each grid until returning to the user's starting grid, so as to obtain the best path for each user.