Near-field passive sensing and fast path tracking method for tourists in scenic spot

Abstract

The invention discloses a near-field passive sensing and fast path tracking method for tourists in a scenic spot, which adopts a face-to-overall feature two-step iteration method to realize cascading of face and body features of a target tourist so as to realize multi-feature coupling re-identification. And then high-precision automatic identification, dynamic accurate tracking and space-time trajectory drawing of the target tourist are realized by combining a video positioning algorithm. Through coupling of body and facial features, the problem that tourists cannot be recognized due to facial feature deficiency can be solved, the acquisition amount of effective identity information is improved, the recognition precision and robustness of an algorithm are remarkably improved, the problems of low tourist target recognition precision, key information loss, path tracking difficulty and the like can be effectively solved, and the tourist target recognition efficiency is improved. Technical support and data support can be provided for safety management of tourists in the scenic area, and the method is particularly suitable for carrying out accurate identity recognition and path tracking on target tourists with weak active positioning signals or non-cooperative target tourists which do not cooperate with safety management in the mountain type open scenic area.

Description

Technical Field

[0001] The present invention relates to a personnel perception and fast path tracking method, specifically a near-field passive perception and fast path tracking method for tourists in mountainous natural scenic areas, belonging to the technical field of intelligent management of tourist safety in scenic areas. Background Art

[0002] Mountainous open scenic areas refer to scenic areas that mainly take natural mountain landscapes as the main tourist objects and provide various tourism activities such as sightseeing, leisure, and exploration, such as Yimeng Mountain Scenic Area, Taishan Scenic Area, etc. Mountainous open scenic areas not only have rich natural landscapes but also contain profound historical and cultural values. The safety issue of tourists in mountainous open scenic areas has always been a key concern for the tourism industry and scenic area management departments. The accurate acquisition and dynamic tracking of the spatio-temporal information of target tourists are of great significance and value for the safety management and emergency rescue of scenic areas. For example: the rapid spatial search and path analysis of special-needs tourists, rapid locking and dynamic tracking, and the clearing and safety management of scenic areas in case of special events / emergency situations. Currently, most popular tourist scenic areas have deployed a large number of automatic cameras and constructed a ground near-field perception network, which can achieve automatic video recording and full coverage of the scenic area. At the same time, it can also automatically identify whether there are people in the image, but there are still difficulties in identifying the identities of tourists. Especially in mountainous open scenic areas, the dynamic perception of small target tourists at a long distance can only rely on post-processing by manual interpretation. Therefore, when conducting activities such as search and rescue in mountainous open scenic areas with a large coverage area, operators still need to analyze the video streams collected by devices such as drones or near-field perception cameras frame by frame. On the one hand, the workload is huge and the efficiency is low. On the other hand, since the search object is relatively small in the environment and usually only occupies a few pixels on the screen, even trained personnel will find it difficult to maintain long-term concentration when looking for individuals in vast mountainous areas or areas with dense vegetation. This will lead to problems such as insufficient real-time acquisition of tourist information in scenic areas and lag in the discovery and handling of safety accidents, seriously restricting the intelligent level of intelligent management of scenic areas.

[0003] The primary challenge faced by the real-time passive perception technology of tourists' spatio-temporal information is how to accurately obtain reliable identity information based on facial features, especially through face recognition technology. Currently, most face recognition technologies are based on machine deep learning, mainly using shallow technologies to learn discriminative image filters, such as DeepFace and FaceNet, training deep neural network (DNN) models for face recognition on large-scale datasets to evaluate face similarity, and achieving relatively high recognition accuracy. However, when applying these face recognition technology models to mountainous open scenic areas, the following problems usually exist: 1. Blurred pixel features of tourists' faces caused by small targets at a distance and poor lighting conditions (for example, blurred facial pixels caused by overexposure); 2. Lack of frontal features caused by different postures of tourists such as lowering the head, turning around, or showing the side (that is, no frontal face features); 3. In complex scenarios, line-of-sight obstacles and decorations may appear during the movement of the target, resulting in the face being blocked (for example, leaves, hats, sunglasses, etc.).

[0004] In response to the above problems, scholars at home and abroad have carried out explorations and research on the passive perception of tourists' spatio-temporal information based on the near-field perception network, and made corresponding improvements to face recognition technology.

[0005] To solve the problem of face blurring caused by small targets at a distance, lighting, etc., one idea for an improved framework is to introduce a margin into the existing loss function to maximize the separation between classes, thereby improving the accuracy and robustness of face recognition. Representative methods include Arcface, Cosface, and SphereFace. Arcface increases the angular distance between classes in the feature space by adding a fixed margin to the angle, thus enhancing the separation between classes; Cosface adds a fixed margin to the cosine similarity between the feature vector and the class weight vector, further widening the distance between different classes, thereby improving the discriminability of features; SphereFace adopts an angular Softmax loss, using the angular margin to enhance the distance between classes.

[0006] To solve the problem of insufficient frontal face features caused by multi-postures such as turning around and showing the side, as well as partial occlusion of the target, X Zhang et al. proposed using the data association results of the target appearance features to judge whether there is occlusion, and reducing the matching error by improving the data association formula; K Cao et al. proposed the Pose-invariant Embedding Network (PENet), which can better handle the impact of posture changes such as turning around and showing the side on face recognition by decoupling the pose features and identity features; J Deng et al. introduced subclasses into the original Arcface to improve its robustness on noisy training data, further enhancing the discriminative ability of features, and showing better adaptability to pose changes and partial occlusion problems.

[0007] However, for the above-improved face recognition technology, its near-field sensing algorithm relying on distributed cameras is mainly applicable to simple environments such as indoor monitoring, schools, or industrial parks. In the complex mountainous open scenic area environment, accurate identification and path tracking of tourists usually face problems such as small targets at a long distance, different postures, and motion occlusion, resulting in insufficient facial features of tourist targets and incomplete features of tourist recognition models. Summary of the Invention

[0008] In view of the problems existing in the above-mentioned prior art, the present invention provides a method for near-field passive perception and fast path tracking of scenic area tourists, which can effectively solve the problems of low accuracy of tourist target recognition, loss of key information, and difficult path tracking in the complex mountainous open scenic area environment, and can provide technical support and data support for the safety management of scenic area tourists. It is particularly suitable for accurately identifying the identities and tracking the paths of target tourists with weak active positioning signals or non-cooperative target tourists who do not cooperate with safety management in mountainous open scenic areas.

[0009] To achieve the above object, the method for near-field passive perception and fast path tracking of scenic area tourists specifically includes the following steps:

[0010] Step1, extract the facial and body features of target tourists in the real-time video stream through the RetinaFace model and the YOLOX model respectively, and apply the ArcFace model to extract the facial features of target tourists, and associate them with the facial features detected by the RetinaFace model;

[0011] Step2, construct a pedestrian re-identification architecture with cascaded coupling of the body and facial features of target tourists, use the Hungarian algorithm to cascade the facial feature and body feature data sets obtained in Step1, construct a multi-feature database, retain the results with iou≥95%, automatically match the face and body of the same target tourist, obtain a real-time cascaded feature set with accurate matching of the face and body of the same target tourist, and call the pedestrian re-identification model trained based on Fast-ReID to re-identify the target tourist;

[0012] Step3, use a video positioning algorithm to map the recognition frame coordinate points to the real world, and complete the trajectory drawing including the spatio-temporal information of target tourists by fitting the spatio-temporal position point set of passive perception of target tourists.

[0013] Furthermore, the specific process of Step1 is as follows:

[0014] ① Based on the Arcface face recognition model, use MTCNN to locate face landmarks for alignment;

[0015] ②Input the backbone network IResnet50 to extract image features, output them to the Head prediction end for feature calculation, obtain the weighted sum logits of each feature after calculating the loss, and extract the facial features of tourists;

[0016] ③Use the Retinaface face detection model. First, input the image, then complete the detection of all tourists' faces in the video stream, and then call the Arcface comparison model to lock the faces of the target tourists in the video frame according to the facial features, and obtain the coordinates of the face recognition frame;

[0017] ④Use the YOLOX detection model, only using the person detection category. Decompose the image frame by frame according to the real-time streaming data, then input it into the backbone network Darknet53 for feature extraction, and finally input it into PAFPN;

[0018] ⑤Use upsampling to complete feature fusion and then downsampling to obtain the feature map, then input it into DecoupledHead for result prediction, correct the prediction box, and obtain the real-time appearance of the target tourist in the picture and the coordinates of the detection box after calculating the loss.

[0019] Furthermore, when extracting the body features of the target tourist in the real-time video stream through the YOLOX model, the YOLOX model is specially trained, the SOD dataset is used, and the dataset of small target tourist pictures in the scenic area is supplemented.

[0020] Furthermore, the specific process of Step2 is as follows:

[0021] ①Read the video stream frame by frame, call Arcface to read the facial pictures provided by the target tourist and compare them with the video frames respectively. For the successfully matched results, input them into the feature set to supplement and strengthen the facial features, and the video frames that fail to be recognized will be resized and then input into Resnet34 for feature reading;

[0022] ②Aggregate the features generated by the backbone network into a global feature through the aggregation module, and then input it into the Head module for normalization and dimensionality reduction processing. The processed features enter the Distance metric stage to compare with the target features in the feature set, so that the body features replace the facial features and complete the confirmation of identity, realizing the coupling from Face to Body;

[0023] ③Adopt the DSR local matching method to generate more detailed local features;

[0024] ④Evaluate and visualize the results, and collect the frames before and after recognition to strengthen the storage of the target appearance features.

[0025] Further, in step ③ of Step 2, after generating more detailed local features using the DSR local matching method, the results are retrieved using two re-ranking methods: K-reciprocal and Query Expansion.

[0026] Further, Step 3 is specifically as follows:

[0027] ① Taking the center of the target tourist's two feet as the initial value of the trajectory node, assuming that the detection box result of the target tourist trajectory node O i in the current frame t m is (u, v, γ, h), where u, v, γ, and h respectively correspond to the abscissa, ordinate, width-to-height ratio, and height of the detection box, then the trajectory node T i of the target tourist in frame t m in the image can be obtained by the following formula: j (u', v') can be obtained by the following formula

[0028]

[0029] In the formula: u is the abscissa of the detection box; v is the ordinate of the detection box; ω is the width of the detection box; δ is the correction value of the target horizontal position; σ is the correction value of the target vertical position;

[0030] ② Connect all the trajectory nodes T j to obtain the tracking trajectory T raj of the target tourist. After fitting all the trajectory nodes, a smooth tracking trajectory is obtained, and the following cubic polynomial is used for trajectory fitting:

[0031] y = a0 + a1x + a2x 2 + a3x 3

[0032] In the formula: a0 is the y-axis intercept coefficient; a1 is the inclination coefficient of the curve; a2 is the parabola shape coefficient; a3 is the cubic nonlinear coefficient.

[0033] Compared with the prior art, the method for near-field passive perception and fast path tracking of tourists in this scenic area uses a two-step iterative method of facial to overall features (F2B: Face to Body two step-wised) to cascade the facial and body features of target tourists, thereby realizing the re-identification of multi-feature coupling. Then, combined with the video positioning algorithm, it realizes the high-precision automatic identification, dynamic precise tracking and spatio-temporal trajectory drawing of target tourists. Experiments show that the overall recognition accuracy of F2B can reach 89.0%, which is 57.4% higher than that of traditional face recognition models. At the same time, the false alarm rate is only 4.8%. Especially when there is a lot of missing target facial feature information, such as postures like lowering the head and turning around, the face is blurred due to high exposure, blocked, and the distance from the camera is far, F2B can solve the problem of unable to identify tourists due to missing facial features and improve the acquisition of effective identity information through the coupling of body and facial features, thereby significantly improving the recognition accuracy and robustness of the algorithm. It can effectively solve the problems of low recognition accuracy of tourist targets, loss of key information, and difficult path tracking in the complex mountainous open scenic area environment, and can provide technical support and data support for the safety management of scenic area tourists. It is especially suitable for accurately identifying the identity and tracking the path of target tourists with weak active positioning signals or non-cooperative target tourists who do not cooperate with safety management in mountainous open scenic areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is the satellite positioning diagram of the experimental area of the embodiment of the present invention;

[0035] Figure 2 is the flow chart of the experiment on near-field passive perception and fast path tracking of tourists in the experimental area of the embodiment of the present invention;

[0036] Figure 3 is the network framework diagram of the F2B model of the present invention;

[0037] Figure 4 is the network framework diagram for obtaining real-time features of target tourists of the present invention;

[0038] Figure 5 is the overall architecture diagram of the pedestrian re-identification model of the present invention;

[0039] Figure 6 is the diagram showing the recognition results of target tourists in the embodiment of the present invention;

[0040] Figure 7 is the diagram showing the spatial positioning and passive path tracking results of target tourists in the embodiment of the present invention;

[0041] Figure 8It is a diagram showing the recognition results of the F2B model of the present invention in a comparative test with the three commonly used models, Facenet, Cosface, and ArcFace, in the scenario of blurred facial features of small targets at a long distance.

[0042] Figure 9 It is a diagram showing the recognition results of the F2B model of the present invention in a comparative test with the three commonly used models, Facenet, Cosface, and ArcFace, in the scenario of multiple angles of tourists' faces.

[0043] Figure 10 It is a diagram showing the recognition results of the F2B model of the present invention in a comparative test with the three commonly used models, Facenet, Cosface, and ArcFace, in the scenario of tourists' faces being occluded. Specific implementation manners

[0044] The method for passive near-field perception and fast path tracking of tourists in this scenic area adopts the two-step iterative method of facial to overall features (F2B: Face to Body two step-wised) as shown in Figure 3 . First, use the near-field video perception network to detect the faces and bodies of tourists by real-time extracting the RTSP video streams of distributed cameras in the scenic area. Secondly, extract local features from facial pictures to real-time lock non-cooperative target tourists, and extract overall features to complete the cascade of facial and body features of target tourists, so as to realize the re-identification of multi-feature coupling. Finally, combine the video positioning algorithm to achieve high-precision automatic recognition, dynamic precise tracking and spatio-temporal trajectory drawing of non-cooperative target tourists.

[0045] Taking the experiment of passive near-field perception and fast path tracking of tourists in Yimengshan Yinzuo Tianmengshan Scenic Area, Feixian County, Linyi City, Shandong Province as an example, the present invention will be further described in conjunction with the accompanying drawings.

[0046] As shown in Figure 1 , the total area of this scenic area is about 240 square kilometers, and the planned core area is about 96 square kilometers. Its terrain is complex, including mountains, forests, waterfalls and red culture. It is famous for its steep terrain and high forest coverage rate, and is a typical mountainous open tourist scenic area. This area is large, the environment is complex, the tree cover is serious, the satellite navigation signal is blocked, and it is difficult to collect information by near-field perception cameras. The camera models equipped in this scenic area are Hikvision DS-2CD2T87E(D)WDA3-L(S) with a maximum resolution of 3840×2160@20fps, and Hikvision DS-2DE7430IW(T5) with a maximum resolution of 2560×1440@25fps. The "test site" area in the figure is the algorithm test and verification point.

[0047] The experimental data was collected on July 4, 2024. A total of 2,274 small-target pictures of scenic spots were collected for supplementing the training of the YoloX pedestrian detection model. The accuracy test data was collected from 9 experimental points, and 100 pictures were collected for each of the scenarios of small targets, heads down, sides, and facial occlusion for accuracy testing. The active Beidou path was collected from the Beidou system positioning terminal, and the actual longitude and latitude coordinates were obtained through calibration at the test points.

[0048] Based on the method of near-field passive perception and fast path tracking of tourists in this scenic area, the entire process of near-field passive perception and fast path tracking experiments of tourists in this scenic area is as Figure 2 shown, and specifically includes the following steps:

[0049] Step1, extract the facial and body features of tourists in the real-time video stream through the RetinaFace model and the YOLOX model respectively, and apply the ArcFace model to extract the facial features of tourists, and associate them with the facial features detected by the RetinaFace model.

[0050] In the extraction of tourists' facial features, considering the superior performance of IResNet50 in the field of facial recognition, ArcFace is selected as the loss function, and IResNet50 is selected as the backbone for feature extraction to improve the accuracy of facial feature extraction. At the same time, the efficient single-stage face detector RetinaFace is combined to detect the faces of tourists in the video stream, enabling it to capture and learn more complex features and adapt to high-precision tasks. RetinaFace is more suitable for application in mountainous open tourist scenic areas with dense faces and complex monitoring environments through a multi-task learning method that combines additional supervision (such as face detection and key point localization) and self-supervision (such as dense key point detection). In addition, a pixel-by-pixel face localization strategy can significantly improve the face detection ability under complex conditions such as large rotations, partial occlusions, long distances, and blurs.

[0051] In the extraction of tourists' body features, a specialized detection model of YOLOX is adopted, and the RetinaFace is used to collect the real-time appearance feature set of the target. YOLOX adopts an anchor-free method and decouples the head from the leading label assignment strategy SimOTA, enabling YOLOX to achieve a better balance between speed and accuracy than other similar products in all model sizes, and is particularly suitable for detecting small targets in long-distance and complex scenarios. Considering the particularity of mountainous open tourist scenic areas, the detection model of YOLOX is specially trained using a small-target dataset (SOD salient object detection dataset), and a dataset containing 2,274 small-target tourist pictures of the study area is supplemented to strengthen the detector's ability to capture small targets at long distances in complex scenarios and meet the detection requirements of tourists in long-distance scenarios in complex mountainous open tourist scenic areas.

[0052] As shown in Table 1 below, the common datasets for training the four models (RetinaFace model, YOLOX model, ArcFace model, Fast-ReID model) for constructing the algorithm mainly come from MS-Celeb-1M_Align_112x112, WIDERFACE, MOT17, Market1501, DukeMTMC-reID, and MSMT17.

[0053] Table 1 Sources of Common Datasets

[0054]

[0055] The acquisition of real-time features of target tourists is based on the reading and comparison of tourists' facial features, as well as the association after real-time face and body detection, as Figure 4 shown, and the specific operation process is as follows:

[0056] ① Based on the classic Arcface face recognition model, with the input image size of 112×112, use MTCNN to locate the face landmarks for alignment.

[0057] ② Input the backbone network IResnet50, extract the image features, output to the Head prediction end for feature calculation, calculate the loss, and obtain the weighted sum logits of each feature, thereby extracting the facial features of special tourists.

[0058] ③ Use the Retinaface face detection model. First, adjust the image size to 640×640 for input, then complete the detection of all tourists' faces in the video stream, and then call the comparison model Arcface to lock the faces of target tourists in the frame according to the facial features, and obtain the coordinates of the face recognition frame.

[0059] ④ Use the specialized YOLOX detection model, only using the detection category of person. Decompose the image into 640×640×3 frame by frame according to the real-time streaming data, input it into the backbone network Darknet53 for feature extraction, and then input it into PAFPN.

[0060] ⑤ Use upsampling to complete feature fusion and then downsampling to obtain the feature map. Immediately input it into DecoupledHead for result prediction, correct the prediction box, calculate the loss, and obtain the real-time appearance and detection box coordinates of tourists and pedestrians in the picture.

[0061] Step 2. To solve the problem of obtaining the appearance features of non-cooperative targets and improve the recognition efficiency, a pedestrian re-identification architecture with cascaded coupling of body and facial features is constructed. The Hungarian algorithm is used to cascade the facial feature and body feature datasets obtained in Step 1, build a multi-feature database, retain the results with an intersection over union (IoU) ≥ 95%, automatically match the face and body of the same tourist, obtain a real-time cascaded feature set with precise matching of the facial and body features of the target tourist, and when searching for the target tourist, call the pedestrian re-identification model trained based on Fast-ReID to re-identify the target tourist.

[0062] Compared with other algorithms for solving the maximum bipartite matching problem, the high efficiency, stability, and accuracy of the Hungarian algorithm enable it to more effectively complete the task of feature cascading. In the tourist re-identification module, a re-identification model trained using the Fast-ReID framework is selected to extract the features of the collected video stream and match them with the facial and body features in the feature set. In the Fast-ReID architecture, multiple iterations of an aggregation module, a head network, a loss function, a distance metric, and post-processing are adopted to ensure the accuracy of the output results.

[0063] The overall architecture of the pedestrian re-identification model is as Figure 5 shown, and the specific operation process is as follows:

[0064] ① Read the video stream frame by frame, call Arcface to read the facial pictures provided by the target tourist and compare them with the video frames. For the successfully matched results, transfer them to the feature set to supplement and strengthen the facial features, while the video frames that fail to be successfully recognized will be resized to 256×128 and then transferred to Resnet34 for feature reading.

[0065] ② Aggregate the features generated by the backbone network into a global feature through the aggregation module, and then transfer it to the Head module for processing such as normalization and dimensionality reduction. The processed features enter the Distance metric stage to be compared with the target features in the feature set. At this time, the body features replace the facial features to complete the confirmation of identity, realizing the coupling from Face to Body, and further realizing the re-identification of feature coupling that complements the body features in the case of insufficient facial features.

[0066] ③ Adopt the DSR (deep spatial reconstruction) local matching method to generate more detailed local features, and use two re-ranking methods, K-reciprocal and Query Expansion (QE), for the results to further improve the accuracy.

[0067] ④Complete the re-identification of the target tourist under multiple cameras by result evaluation and visual output, and collect the frames before and after the identification to strengthen the storage of the target appearance features.

[0068] The target tourist recognition results of the specific target tourist experiment based on the F2B technology framework in the Yimeng Mountain Yinzuo Tianmeng Mountain Scenic Area are as Figure 6 shown in the figure. In the figure: the blue box is the face recognition box, the red box is the F2B recognition box, "found_person" is the label of the target tourist, and "None / Unknown" is the label of the unrecognized tourist. Figure 6 (A) and Figure 6 (B) respectively show the video frames of the side of the target tourist and the recognition results in the situation of looking down. Figure 6 (C) shows the recognition results when part of the facial features of the target tourist are blocked. Figure 6 (D) is the recognition result in the long-distance scene. The target tourist is about 50 meters away from the camera, and there are only 20 pixels for the facial features, which are difficult to identify and the facial feature information is seriously insufficient. It can be seen that the F2B method can effectively complete the re-identification of specific target tourists in various scenarios and postures, thus greatly improving the effective recognition rate of target tourists.

[0069] Step3, use the video positioning algorithm to map the coordinate points of the recognition box to the real world, and complete the trajectory drawing including the spatio-temporal information of the target tourist by fitting the spatio-temporal position point set of the target tourist's passive perception.

[0070] Adopt the video positioning algorithm that fuses the dynamic target in the surveillance video with the geospatial information. Take the center of the target tourist's two feet as the initial value of the trajectory node, and then perform trajectory correction according to the pose of the camera relative to the ground plane and the proportion of the target in the image.

[0071] Suppose the detection box result of a certain tracking target O i in the current frame t m obtained from the multi-target tracker is (u, v, γ, h), corresponding to the horizontal and vertical coordinates of the lower left point of the detection box, the width-to-height ratio, and the height respectively. Then the trajectory node T i of the target O m in the t j frame in the image can be obtained by the following formula

[0072]

[0073] In the formula: u is the abscissa of the detection box; v is the ordinate of the detection box; ω is the width of the detection box; δ is the correction value of the target's horizontal position; σ is the correction value of the target's vertical position.

[0074] Since the tracking target O obtained above mTrajectory node T j There is an error, so connect the trajectory node T j The resulting trajectory T raj When jitter occurs, all trajectory nodes need to be fitted to obtain a smooth tracking trajectory. This is done using a cubic polynomial, as shown in the following equation. The coefficients of a cubic polynomial can be expressed as a matrix, called the Hessian matrix, which describes the curvature and shape of a function at a specific point.

[0075] y=a0+a1x+a2x 2 +a3x 3

[0076] Where: a0 is the y-axis intercept coefficient, which determines the starting height of the curve; a1 is the slope coefficient of the curve, which controls the trend of linear change; a2 is the parabola shape coefficient, which is used to control the curvature of the curve; a3 is the cubic nonlinear coefficient, which is used to capture more complex nonlinear changes and increase the flexibility of the trajectory.

[0077] The sum of squared deviations of each node to the tracking trajectory is as follows:

[0078]

[0079] Where: y i is the actual y value of the i-th trajectory node, that is, the vertical coordinate of the target image in the i-th frame; a0+a1x+a2x 2 +a3x 3 The fitting curve is at the i-th node x i The corresponding y value.

[0080] After taking partial derivatives, we get the mapping matrix expression shown below

[0081]

[0082] Right now

[0083] X×A=Y

[0084] A=(X T X) -1 X T Y

[0085] Where: X is an n×(k+1) matrix, called the Vandermonde matrix, and each row corresponds to the x of the trajectory node iThe polynomial form of value expansion; A is a (k + 1)×1 column vector representing the coefficients of the fitting polynomial, which is the final solution, the coefficients of each order of the fitting curve, and minimizes the sum of the squared deviations between the trajectory nodes and the fitting curve; Y is an n×1 column vector representing the actual y-values of the trajectory nodes, that is, the vertical coordinates of the target in each frame; (X T X) -1 is used to standardize or scale the trajectory nodes to ensure that the final solution is unique and optimal; X T Y represents the correlation between the actual y-values of the trajectory nodes and the corresponding x-values.

[0086] Then the set of trajectory nodes of the tracking target O m in the image space is represented as follows:

[0087]

[0088] In the formula: u' j represents the horizontal coordinate of each node; v' j represents the vertical coordinate of each node; each column of the matrix represents a trajectory node, indicating the spatial coordinate information of the target at a certain time point.

[0089] The set of smooth trajectory nodes in the geospatial obtained after the above 3-degree polynomial fitting is represented as follows:

[0090]

[0091] In the formula: represents the coordinates of the j-th trajectory node in the geospatial; H -1 represents the inverse transformation matrix of the perspective transformation, which converts the image coordinates back to the geospatial coordinates; K is the internal parameter matrix of the camera, containing information such as the focal length and the optical center; [r1, r2|t] is the external parameter matrix, where r1 and r2 are the rotation matrices of the camera, describing the rotation of the camera relative to the world coordinate system; t is the translation vector of the camera, describing the position of the camera in the world coordinate system; A' is a vector with extended dimensions, related to homogeneous coordinates and perspective transformation, and is used for spatial geometric transformation.

[0092] In the scene of Yimeng Mountain Yinzuo Tianmeng Mountain Scenic Area, the results of the spatial positioning and passive path tracking of the target tourists in the specific target tourist experiment based on the F2B technology framework are as Figure 7 shown, Figure 7 (A) shows the passive perception results of the accurate spatio-temporal information of the target tourists and the tracking of the path; Figure 7(B) shows the comparison between the passive perception path of the target tourist and the active perception positioning result. The passive perception result of the detailed spatio-temporal information of the target tourist based on F2B is shown in Table 2. It can be seen from Table 2 that the video frames of each camera can return the recognized tourist coordinate points 2 to 4 times, and the integrity of the trajectory collection reaches 100%. The gap with the actively collected trajectory is very small, with high credibility.

[0093] Table 2 Passive perception results of the detailed spatio-temporal information of the target tourist

[0094]

[0095]

[0096] To verify the superiority of the F2B algorithm in different scenarios, three commonly used models, Facenet, Cosface, and ArcFace, were selected for comparative testing. In the test, the threshold of F2B was set to 0.6. The selected target tourist was labeled with "found_person", and the rest of the tourists were labeled with "Unknown / None". The comparative test results are as Figure 8 , Figure 9 , Figure 10 shown.

[0097] Figure 8 shows the recognition comparison results in the scenario of blurred facial features of small targets at a long distance. It can be seen from a-1 to a-4 and c-1 to c-4 in the figure that at a long distance, F2B and the three comparison models perform very well in the detection ability of small targets, but only F2B successfully completed the recognition of small targets at a long distance. As can be seen from b-1 to b-4 in the figure, due to the lighting reason, the face of the tourist shows a high-exposure phenomenon, and only F2B successfully recognized the identity of the target tourist. Thus, in scenarios such as blurred facial pixels at a long distance, F2B can effectively complete the recognition of the identity of the target tourist by supplementing body features.

[0098] Figure 9 shows the comparative test results of tourists in the scenario of multi-angle faces. The main problem faced in this scenario is the lack of frontal face features caused by different postures of tourists such as lowering their heads, turning around, and being in the side position. a-1 to a-4 in the figure show that the target tourist is in a head-down state, and only F2B completed the recognition of the target tourist. b-1 to b-4 in the figure show that the target tourist is in a back state, and only very few facial features can be captured. Although F2B and the other three comparison models all completed the detection of the target tourist, only F2B achieved the recognition of the identity of the target tourist. c-1 to c-4 in the figure show that the target tourist is in a side state, and at this time, F2B and the three comparison models all completed the recognition of the target tourist. Thus, F2B has more advantages when the tourist is in the back or head-down state with very few facial features.

[0099] Figure 10 Shows the comparative test results in the face occlusion scenario, that is, the face occlusion problem caused by the existence of line-of-sight obstacles and decorations (such as leaves, hats, sunglasses, etc.) during the movement of the target tourist. It can be seen from a-1 to a-4, b-1 to b-4, and c-1 to c-4 in the figure that the target tourist is wearing white clothes and a gray mask, and the face is severely occluded. Therefore, Facenet, Cosface, and Arcface cannot effectively identify the identity of the target tourist. Only F2B can successfully capture the identity information relying on body features under such harsh conditions and complete the identification of the target tourist.

[0100] To further test the accuracy and stability of the F2B algorithm, three commonly used models, Facenet, Cosface, and ArcFace, were also selected for comparative testing, and quantitative tests were carried out for each scenario. Video frames were randomly selected from scenarios such as long-distance small targets, different postures, and face occlusion. 100 frames were taken for each scenario, forming a total of 400 test and validation data sets. And a quantitative statistical analysis was carried out on the overall accuracy of the results. The statistical results are shown in Table 3. It can be seen from Table 3 that F2B far exceeds the three comparison models in terms of overall accuracy. Especially when the target is at a long distance, with the head down, and face occlusion, the recognition success rates reach 80%, 95%, and 85% respectively, showing a significant improvement. When the target is in the side state, since more facial features can be collected at this angle, the difference in the recognition success rate is smaller. It can be seen that the results of the F2B algorithm are more stable and reliable.

[0101] Table 3 Comparison of recognition accuracies of different models

[0102]

[0103] To further verify the stability and reliability of F2B, three commonly used models, Facenet, Cosface, and ArcFace, were also selected for comparative testing, and a statistical analysis of the false alarm rate, a quantitative index, was carried out respectively. The statistical results are shown in Table 4. It can be seen from Table 4 that the overall false alarm rate of the F2B algorithm is less than 5%, and the credibility of its recognition results is the highest. However, for long-distance small targets, the false alarm rate of the F2B algorithm is higher than that of Facenet, Cosface, and Arcface. This is because the number of successful identifications of the F2B in small target recognition is much larger than that of other comparison models. The F2B successfully identified 80 out of 100 experimental pictures. In this scenario, the false alarm rate of F2B is only 10%, while the false alarm rates of Facenet, Cosface, and Arcface are relatively low, mainly because the total number of identifications of Facenet and Arcface is 0, and only Cosface successfully identified one.

[0104] Table 4 Comparison of False Alarm Rates Identified by Different Models

[0105]

[0106] The comparison results of F2B and other face recognition models in the three common scenarios of mountainous open scenic spots show that the overall recognition accuracy of F2B can reach 89.0%, which is 57.4% higher than that of traditional face recognition models. At the same time, the false alarm rate is only 4.8%. Especially when there is a lot of missing target facial feature information, such as postures like lowering the head and turning around, facial blurring due to high exposure, occlusion, and being far away from the camera, F2B can solve the problem of unable to recognize tourists due to missing facial features and improve the acquisition of effective identity information through the coupling of body and facial features, thus significantly improving the recognition accuracy and robustness of the algorithm. In the complex environment of mountainous open scenic spots, it can effectively solve problems such as low recognition accuracy of tourist targets, loss of key information, and difficulty in path tracking, and can provide technical support and data support for the safety management of scenic spot tourists. It is especially suitable for accurately identifying the identity and tracking the path of target tourists with weak active positioning signals or non-cooperative target tourists who do not cooperate with safety management in mountainous open scenic spots.

Claims

1. A method for near-field passive perception and rapid path tracking of scenic area tourists, characterized in that, Specifically, it includes the following steps: Step1: Extract the facial and body features of the target tourist in the real-time video stream through the RetinaFace model and the YOLOX model respectively, and apply the ArcFace model to extract the facial features of the target tourist, and associate them with the facial features detected by the RetinaFace model; Step2: Construct a person re-identification architecture with cascaded coupling of the target tourist's body and facial features. Use the Hungarian algorithm to cascade the facial feature and body feature datasets obtained in Step1, construct a multi-feature database, retain the results with iou≥95%, automatically match the face and body of the same target tourist, obtain a real-time cascaded feature set with accurate matching of the face and body of the same target tourist, and call the person re-identification model trained based on Fast-ReID to re-identify the target tourist; Step3: Use a video localization algorithm to map the coordinates of the recognition box to the real world, and complete the trajectory drawing including the spatio-temporal information of the target tourist by fitting the set of spatio-temporal position points passively perceived by the target tourist.

2. The method for near-field passive perception and fast path tracking of scenic area visitors according to claim 1, wherein The specific process of Step1 is as follows: ① Based on the Arcface face recognition model, use MTCNN to locate the face landmarks for alignment; ② Input the backbone network IResnet50, extract image features, output to the Head prediction end for feature calculation, calculate the loss to obtain the weighted sum logits of each feature, and extract the facial features of the tourist; ③ Use the Retinaface face detection model. First, input the image, then complete the detection of the faces of all tourists in the video stream, and then call the Arcface comparison model to lock the face of the target tourist in the frame according to the facial features, and obtain the coordinates of the facial recognition box; ④ Use the YOLOX detection model, only use the detection category of person, decompose the image frame by frame according to the real-time streaming data, then input it into the backbone network Darknet53 for feature extraction, and finally input it into PAFPN; ⑤ Use upsampling to complete feature fusion and then downsampling to obtain a feature map, then input it into Decoupled Head for result prediction, and correct the prediction box, and calculate the loss to obtain the real-time appearance of the target tourist in the picture and the coordinates of the detection box.

3. The scenic area visitor near-field passive perception and fast path tracking method according to claim 1 or 2, characterized in that, When extracting the body features of the target tourist in the real-time video stream through the YOLOX model, the YOLOX model is specially trained, the SOD dataset is used, and the dataset of pictures of small target tourists in the scenic area is supplemented.

4. The method for near-field passive perception and fast path tracking of scenic area visitors according to claim 1, characterized in that The specific process of Step2 is as follows: ① Read the video stream frame by frame, call Arcface to read the facial pictures provided by the target tourist and compare the video frames. For the successfully matched results, pass them into the feature set to supplement and strengthen the facial features, and the video frames that fail to be successfully recognized will be resized and then passed into Resnet34 for feature reading; ②Aggregate the features generated by the backbone network into a global feature through the aggregation module, and then pass it into the Head module for normalization and dimensionality reduction processing. After processing, the features enter the Distance metric stage to be compared with the target features in the feature set, so that the body features replace the facial features to complete the identity confirmation and achieve the coupling from Face to Body; ③Adopt the DSR local matching method to generate more detailed local features; ④Evaluate and visualize the results, and collect the frames before and after recognition to strengthen the storage of the target appearance features.

5. The scenic area visitor near-field passive perception and fast path tracking method according to claim 4, characterized in that In step ③, after adopting the DSR local matching method to generate more detailed local features, the K-reciprocal and Query Expansion re-ranking methods are used to retrieve the results.

6. The method for near-field passive perception and fast path tracking of scenic area visitors according to claim 1, characterized in that Step 3 is specifically as follows: ①Taking the center of the target tourist's feet as the initial value of the trajectory node, assuming the current frame t obtained i The detection box result of the target tourist trajectory node O m in is (u, v, γ, h), where u, v, γ, h correspond to the abscissa, ordinate, width-to-height ratio, and height of the detection box respectively, then t i The target tourist trajectory node O m The trajectory node T of in the image j (u', v') can be obtained by the following formula In the formula: u is the abscissa of the detection box; v is the ordinate of the detection box; ω is the width of the detection box; δ is the correction value of the target horizontal position; σ is the correction value of the target vertical position; ②Connect all the trajectory nodes T j Obtain the tracking trajectory T of the target tourist raj , after fitting all the trajectory nodes, obtain a smooth tracking trajectory, and use the cubic polynomial shown in the following formula for trajectory fitting: y = a0 + a1x + a2x 2 + a3x 3 In the formula: a0 is the y-axis intercept coefficient; a1 is the inclination coefficient of the curve; a2 is the parabola shape coefficient; a3 is the cubic nonlinear coefficient.

Images