Police unmanned aerial vehicle prevention and control dispatching system

Abstract

The application provides a grassroots police unmanned aerial vehicle prevention and control scheduling system, and belongs to the field of prevention and control scheduling. The system comprises the following steps: obtaining a first transition feature by acquiring a pre-blind area track and a blind area entrance feature; performing feature correlation degree calculation on the first transition feature and the blind area entrance feature; when the feature correlation degree is greater than a preset threshold, a blind area crossing guide path is determined, and the police unmanned aerial vehicle is controlled to fly to a monitoring blind area according to the blind area crossing guide path, and a blind area exit feature behind the monitoring blind area is acquired; the blind area exit feature is reversely associated with the blind area entrance feature and the first transition feature, and if the reverse association is successful, a continuation feature of a suspicious target is determined; and a complete track of the suspicious target in the monitoring blind area is determined by the continuation feature and the pre-blind area track, and a tracking prevention and control route of the police unmanned aerial vehicle is corrected. The technical scheme provided by the application can realize accurate prevention and control tracking of a suspicious target in a monitoring blind area by a police unmanned aerial vehicle under a complex monitoring environment.

Description

Technical Field

[0001] This application relates to the field of prevention and control dispatch technology, and more specifically, to a grassroots police drone prevention and control dispatch system. Background Technology

[0002] The security of large-scale urban events, emergency response, and low-altitude safety control have placed higher demands on police work. Traditional ground patrols have limitations such as limited visibility and slow response speed, making it difficult to meet the needs of three-dimensional prevention and control. The application of police drone technology has enabled the transformation from two-dimensional planar patrols to three-dimensional prevention and control. Through deep integration with the command and dispatch system, an "air-ground integrated" prevention and control dispatch platform has been built. This platform relies on multi-source data fusion and intelligent analysis technology to integrate drone automatic nesting, spectrum detection, countermeasures, and a GIS visualization command platform. It can monitor the low-altitude airspace situation in real time, realize an integrated air-ground security defense line, and greatly shorten the emergency response time.

[0003] Chinese patent CN114756053A relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to a UAV-based police patrol method, system, and storage medium. The method includes the following steps: collecting regional information from the UAV; determining whether a security incident has occurred in the corresponding area based on the regional information; if a security incident exists, sending a tracking request to the corresponding UAV to collect information on the individuals involved; obtaining real-time tracking video information about the individuals involved collected by the UAV; and sending the real-time tracking video information to a command center platform. This application obtains regional information and makes a judgment: if it determines that a security incident exists in the area, it allows the UAV to track the individuals involved in real time; sending the real-time tracking video information to the command center platform helps the command center platform to dispatch police officers to intercept and arrest the individuals involved based on their movement trajectories, thus helping to control security incidents in a timely manner and reducing losses to personnel and property. This solution addresses the issue of "real-time tracking of individuals involved in security incidents," but it presupposes that the presence of an incident can be identified in the area information and that the target remains within the drone's field of view. However, it fails to effectively handle situations where the target's trajectory breaks down after it actively or passively enters a blind spot that neither fixed surveillance nor the drone can cover. When a suspicious target enters an uncovered blind spot from a fixed ground surveillance point, the system cannot effectively utilize the discontinuity before and after the blind spot to identify the target's identity, nor can it accurately guide the drone through the blind spot for continued tracking. This leads to the drone easily misidentifying the target as another similar target when it recaptures it, thereby reducing the collaborative prevention and control accuracy of police drones in complex surveillance environments. Therefore, how to achieve accurate prevention and tracking of suspicious targets in surveillance blind spots by police drones in complex surveillance environments has become a challenge for the industry. Summary of the Invention

[0004] This application provides a grassroots police drone control and dispatch system, which can enable police drones to accurately control and track suspicious targets in blind spots of monitoring in complex monitoring environments.

[0005] This application provides a grassroots police drone control and dispatch system, which includes: The acquisition module is used to receive the trajectory of a suspicious target before entering the blind zone and the characteristics of the blind zone entrance sent by the ground fixed monitoring equipment, and to dispatch a police drone to collect the first transition features with the end point of the trajectory before entering the blind zone as the first connection point. The guidance path determination module is used to calculate the feature correlation degree between the first transition feature and the blind spot entrance feature. When the feature correlation degree is greater than a preset threshold, a blind spot crossing guidance path is determined based on the first connection point and the blind spot front trajectory, and the police drone is controlled to fly to the monitoring blind spot according to the blind spot crossing guidance path. The continuity feature recognition module is used to obtain the blind zone exit feature behind the monitoring blind zone after the police drone flies to the monitoring blind zone, and to reverse associate the blind zone exit feature with the blind zone entrance feature and the first transition feature. If the reverse association is successful, the blind zone exit feature is marked as the continuity feature of the same suspicious target. The flight path correction module is used to perform spatiotemporal extrapolation of the movement trajectory of the suspicious target by using the continuity features and the trajectory before the blind zone, thereby generating the complete trajectory of the suspicious target in the monitoring blind zone, and correcting the tracking and control flight path of the police drone based on the complete trajectory.

[0006] Furthermore, in the acquisition module, receiving the trajectory of a suspicious target before entering the blind zone and the entry characteristics of the blind zone from the ground-based fixed monitoring equipment specifically includes: Motion masks of suspicious targets are extracted from continuous video streams from ground-based fixed monitoring equipment. Target identifiers are assigned to the motion masks using a multi-target tracking algorithm, and trajectory segments with the target identifiers are generated. Determine whether the end point of the trajectory segment enters the preset monitoring blind zone boundary. When it enters the monitoring blind zone boundary, mark all trajectory points of the trajectory segment as the trajectory before the blind zone. Within a preset time window before the end point of the trajectory before the blind zone, depth features are extracted from the image block of the suspected target to generate blind zone entry features.

[0007] Furthermore, in the acquisition module, the process of dispatching the police drone to acquire the first transition feature using the end point of the trajectory before the blind spot as the first connection point specifically includes: Using the end point of the trajectory before the blind zone as the first continuation point, the first continuation point is converted from the pixel coordinate system of the ground fixed monitoring equipment to a unified world coordinate system, generating the three-dimensional spatial coordinates of the first continuation point; A hovering data acquisition waypoint is generated with the three-dimensional spatial coordinates of the first connection point as the center. After the police drone reaches the hovering data collection point, the onboard visual sensor of the police drone projects an image of the entrance area of ​​the monitoring blind spot onto the ground in the direction corresponding to the three-dimensional spatial coordinates of the first connection point, thereby obtaining an image of the entrance area of ​​the blind spot. Suspicious targets in the blind zone entrance area image are identified by feature recognition to generate a first transition feature.

[0008] Furthermore, in the guidance path determination module, the first transition feature and the blind spot entrance feature are used to calculate the feature correlation degree. When the feature correlation degree is greater than a preset threshold, the blind spot crossing guidance path is determined based on the first connection point and the blind spot pre-trajectory, specifically including: Calculate the cosine similarity between the first transition feature and the blind zone entrance feature as the feature correlation degree; When the feature correlation is greater than a preset threshold, it is confirmed that the target corresponding to the first transition feature and the target corresponding to the blind zone entrance feature are the same suspicious target, and the motion direction vector corresponding to the blind zone pre-trajectory is determined. Using the first connection point as the starting point of the path planning and the motion direction vector as the initial crossing direction, a smooth path curve is generated that passes through the monitoring blind zone and ends behind the monitoring blind zone as the blind zone crossing guidance path.

[0009] Furthermore, in the guidance path determination module, controlling the police drone to fly through the blind spot and reach the monitoring blind spot specifically includes: The blind spot traversal guidance path is discretized into a three-dimensional waypoint sequence with timestamps; The timestamped three-dimensional waypoint sequence is loaded into the flight path buffer of the police drone, triggering the police drone to perform a crossing maneuver in the waypoint sequence until it reaches the destination waypoint behind the monitoring blind spot.

[0010] Furthermore, in the continuity feature recognition module, after the police drone flies to the monitoring blind spot, acquiring the blind spot exit feature behind the monitoring blind spot specifically includes: When the police drone reaches the end point of the guidance path through the blind spot, the onboard visual sensor of the police drone is controlled to perform a wide-area scan of the exit area behind the monitoring blind spot, centered on the three-dimensional spatial coordinates of the end point, and to obtain the video stream of the exit area. All moving targets appearing in the video stream of the export area are segmented into instances to generate a contour mask and its corresponding depth feature vector for each moving target. Calculate the feature correlation degree between each depth feature vector and the first transition feature, and filter out the moving target with the highest feature correlation degree as the suspected same suspicious target; Extract the pose key point heatmap, local texture gradient histogram and color space distribution vector of the suspected same suspicious target in the current frame, and then obtain the blind zone exit features behind the monitoring blind zone.

[0011] Furthermore, in the continuity feature recognition module, the blind zone exit feature is reverse-associated with the blind zone entrance feature and the first transition feature. If the reverse association is successful, marking the blind zone exit feature as a continuity feature of the same suspicious target specifically includes: The blind zone exit feature and the blind zone entrance feature are input into the first twin network branch for cross-domain measurement to obtain the first reverse similarity score. At the same time, the blind zone exit feature and the first transition feature are input into the second twin network branch for cross-domain measurement to obtain the second reverse similarity score. The first reverse similarity score and the second reverse similarity score are weighted and fused to generate a reverse association confidence score; If the confidence level of the reverse association exceeds the preset confirmation threshold, the reverse association is determined to be successful, and the blind zone exit feature is marked as a successor feature of the same suspicious target.

[0012] Furthermore, in the flight path correction module, correcting the tracking and control flight path of the police drone based on the completed trajectory specifically includes: The completed trajectory is projected from the world coordinate system onto the onboard real-time local grid map of the police drone to generate a reference path layer containing all spatiotemporal poses of the suspicious target within the monitoring blind zone; Using the reference path layer as a planning constraint, a gradient-based trajectory optimizer generates a smooth tracking and control route that satisfies kinematic constraints and minimizes the average distance to the reference path layer between the current waypoint of the police drone and the latest detection position corresponding to the successive feature.

[0013] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: The grassroots police drone control and dispatch system provided in this application firstly includes a data acquisition module, used to receive the pre-blind zone trajectory and blind zone entrance features of a suspicious target sent by ground-based fixed monitoring equipment. The module then dispatches the police drone to collect the first transition feature using the end point of the pre-blind zone trajectory as the first connection point. Secondly, a guidance path determination module calculates the feature correlation between the first transition feature and the blind zone entrance feature. When the feature correlation is greater than a preset threshold, a blind zone crossing guidance path is determined based on the first connection point and the pre-blind zone trajectory, and the police drone is controlled to fly to the blind zone crossing guidance path. The monitoring blind zone is identified. Then, a follow-up feature recognition module is used to acquire the blind zone exit feature behind the monitoring blind zone after the police drone flies to the monitoring blind zone, and to reverse-associate the blind zone exit feature with the blind zone entrance feature and the first transition feature. If the reverse association is successful, the blind zone exit feature is marked as a follow-up feature of the same suspicious target. Finally, a flight path correction module is used to perform spatiotemporal deduction of the movement trajectory of the suspicious target through the follow-up feature and the trajectory before the blind zone, thereby generating the complete trajectory of the suspicious target in the monitoring blind zone, and correcting the tracking and control flight path of the police drone based on the complete trajectory.

[0014] Therefore, this application demonstrates that police drones can accurately track and control suspicious targets in blind spots under complex monitoring environments. First, the data acquisition module establishes a cross-domain collaborative perception mechanism between fixed ground monitoring and police drones, enabling the drone to automatically reconnect with target cues broken at blind spot boundaries, eliminating information silos between fixed monitoring and mobile tracking, and providing initial spatiotemporally aligned association points for subsequent cross-blind spot tracking. Second, the guidance path determination module generates a crossing path based on the target's movement direction prior by cross-domain feature association between the transition features of the drone's aerial view and the ground entrance features. This allows the drone to actively cross blindly in blind spots where the target is completely invisible, rather than blindly searching or passively hovering, thus minimizing the target loss time window and preventing exit re-capture failure due to incorrect crossing direction. Finally, the continuity feature recognition module generates continuity features using a two-way cross-validation mechanism from the exit to the entrance and transition points, rather than relying solely on one-way feature matching, effectively... This invention solves the identity switching errors caused by sudden changes in perspective and lighting differences in traditional cross-camera tracking, ensuring that the target re-locked behind the blind spot is the same suspicious target as the one in front of the blind spot. This provides reliable identity continuity for subsequent trajectory completion and tracking, avoiding misjudgment of other similar targets when the drone re-captures the target, thereby improving the collaborative prevention and control accuracy of police drones in complex monitoring environments. Finally, the flight path correction module uses continuity features and the trajectory before the blind spot to perform spatiotemporal trajectory extrapolation based on motion pattern learning and geographical constraints, generating a physically reasonable complete trajectory for the target within the blind spot and correcting the drone's tracking flight path accordingly. This allows the drone to obtain a tracking guidance line that closely follows the target's most likely movement path and meets flight safety constraints the moment the target reappears, avoiding the defect of only being able to passively react after the target reappears. This achieves uninterrupted active tracking and control closed loop throughout the entire process across blind spots. In summary, the technical solution provided in this application can enable police drones to accurately prevent and track suspicious targets in monitoring blind spots in complex monitoring environments. Attached Figure Description

[0015] Figure 1 This is an application scenario diagram of the grassroots police drone prevention and control dispatch system shown in this application; Figure 2 This is a module structure diagram of the grassroots police drone prevention and control dispatch system shown in this application; Figure 3 This is an exemplary flowchart illustrating the determination of a blind spot traversal guidance path according to some embodiments of this application. Detailed Implementation

[0016] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] refer to Figure 1 The figure is a schematic diagram of the application scenario of the grassroots police drone prevention and control dispatch system shown in this application. The figure includes a surveillance camera, a control center, and a police drone. The surveillance camera is used to collect and transmit to the control center the trajectory before the suspicious target enters the blind zone and the blind zone entrance features. At the same time, it can send the first connection point information to the police drone. The control center calculates the feature correlation degree based on the blind zone trajectory before the blind zone and the blind zone entrance features transmitted back by the surveillance camera, generates a blind zone crossing guidance path, and issues a dispatch command to the police drone. The police drone receives the command from the surveillance camera and the control center, collects the first transition feature with the end point of the blind zone trajectory as the first connection point, and flies to the monitoring blind zone according to the blind zone crossing guidance path determined by the control center. It obtains the blind zone exit features and transmits them back to the control center. The control center completes the feature reverse correlation and trajectory spatiotemporal extrapolation, generates a complete trajectory, and corrects the tracking and control route to achieve seamless connection and continuous prevention and control of suspicious targets.

[0018] refer to Figure 2 As shown in the figure, this is a modular structure diagram of the grassroots police drone prevention and control dispatch system according to this application. The system includes: a data acquisition module 100, a guidance path determination module 200, a connection feature recognition module 300, and a flight path correction module 400, which are described below: The acquisition module 100 is used to receive the trajectory of a suspicious target before entering the blind zone and the entry features of the blind zone sent by the ground fixed monitoring equipment, and to dispatch a police drone to collect the first transition features with the end point of the trajectory before entering the blind zone as the first connection point.

[0019] It should be noted that, in this application, a suspicious target refers to a moving object that has been initially marked as potentially risky by the ground-based fixed monitoring equipment before entering the blind zone, based on preset behavioral rules or image recognition rules, but cannot be further tracked by the ground-based fixed monitoring equipment after entering the blind zone. Such moving objects include, but are not limited to, people climbing over fences, vehicles driving at low speeds without license plates, and individuals or groups that suddenly change their direction of movement after loitering in key defense areas. These will not be elaborated further here.

[0020] It should also be noted that, in this application, the ground-based fixed monitoring equipment refers to security cameras that are pre-deployed on urban roads, entrances and exits of key areas, perimeters of communities, or the edge of security zones for large events, and have fixed installation locations and fixed monitoring angles, which will not be elaborated further here.

[0021] In the acquisition module, receiving the trajectory of a suspicious target before entering the blind zone and the entry characteristics of the blind zone from the ground-based fixed monitoring equipment is achieved through the following steps: Motion masks of suspicious targets are extracted from continuous video streams from ground-based fixed monitoring equipment. Target identifiers are assigned to the motion masks using a multi-target tracking algorithm, and trajectory segments with the target identifiers are generated. Determine whether the end point of the trajectory segment enters the preset monitoring blind zone boundary. When it enters the monitoring blind zone boundary, mark all trajectory points of the trajectory segment as the trajectory before the blind zone. Within a preset time window before the end point of the trajectory before the blind zone, depth features are extracted from the image block of the suspected target to generate blind zone entry features.

[0022] In specific implementation, firstly, the embedded processing unit built into the ground-based fixed monitoring equipment runs the YOLOv8-seg instance segmentation model to classify each frame of the continuous video stream pixel by pixel, extracting motion masks of suspicious targets. These motion masks refer to the binarized foreground pixel regions obtained through threshold segmentation, used to separate the target from the background. Next, the motion masks are input into the ByteTrack multi-target tracking algorithm. This algorithm uses Kalman filtering to predict the motion state of the bounding box center coordinates, aspect ratio, and area change rate of the target corresponding to each motion mask. It also combines this with the Hungarian algorithm to perform optimal matching by minimizing the cost matrix between the detection box of the current frame and historical trajectory segments. Each continuously appearing motion mask is assigned an integer as a target identifier, thereby generating trajectory segments with the target identifier. These trajectory segments are sequences of continuous target bounding box center point coordinates ordered by timestamps. Then, the coordinates of the bottom midpoint of the latest bounding box in the trajectory segment are compared with the pre-defined... Geometric collision detection is performed on points within the calibrated monitoring blind zone boundary polygon. When the coordinates of the midpoint of the base edge first fall inside the blind zone boundary polygon, it is determined that the point has entered the preset monitoring blind zone boundary. The coordinates of all the center points of the bounding boxes stored in the trajectory segment are taken as trajectory points. The trajectory point sequence, which is combined in chronological order, is marked as the pre-blind zone trajectory. The pre-blind zone trajectory refers to the outer envelope of the uncovered area formed by building obstruction between the field of view of the ground fixed monitoring equipment and the patrol airspace of the police drone. Finally, all image blocks within a preset time window before the end point of the pre-blind zone trajectory are selected. Each image block is scaled to 224×224 pixels and then sequentially input into the ResNet50-IBN-a re-identification backbone network pre-trained on the Market-1501 dataset. The 2048-dimensional feature vector output by the global average pooling layer of the network is taken and then compressed by mean pooling along the time dimension to generate a 2048-dimensional normalized feature vector as the blind zone entry feature.

[0023] It should be noted that the blind zone entry feature in this application refers to the identity discrimination vector obtained by compactly encoding the appearance attributes of a suspicious target before it enters the blind zone. When a suspicious target is about to enter the blind zone, the ground fixed monitoring equipment can collect the last continuous, stable and unobstructed reference feature of the target. This feature serves as the only credible identity anchor point before the target is lost, and can provide a comparison benchmark for the first transition feature collected by the UAV at the first connection point.

[0024] In the data acquisition module, the process of dispatching the police drone to acquire the first transition feature using the end point of the trajectory before the blind spot as the first connection point is implemented through the following steps: Using the end point of the trajectory before the blind zone as the first continuation point, the first continuation point is converted from the pixel coordinate system of the ground fixed monitoring equipment to a unified world coordinate system, generating the three-dimensional spatial coordinates of the first continuation point; A hovering data acquisition waypoint is generated with the three-dimensional spatial coordinates of the first connection point as the center. After the police drone reaches the hovering data collection point, the onboard visual sensor of the police drone projects an image of the entrance area of ​​the monitoring blind spot onto the ground in the direction corresponding to the three-dimensional spatial coordinates of the first connection point, thereby obtaining an image of the entrance area of ​​the blind spot. Suspicious targets in the blind zone entrance area image are identified by feature recognition to generate a first transition feature.

[0025] In specific implementation, firstly, after receiving the blind zone pre-trajectory sent by the ground-based fixed monitoring equipment, the UAV ground control station extracts the end point of the blind zone pre-trajectory, uses the end point as the first connection point, and obtains the corresponding pixel coordinates. Combining the projection equation formed by the factory-calibrated intrinsic parameter matrix of the ground-based fixed monitoring equipment and the extrinsic rotation matrix and translation vector measured by a total station during deployment, the pixel coordinates of the first connection point are converted into three-dimensional spatial coordinates in a unified world coordinate system using the collinear equation back projection method. The world coordinate system refers to the Northeast-Northeast coordinate system with the UAV takeoff point as the origin, east as the X-axis, north as the Y-axis, and the sky direction as the Z-axis. The intrinsic parameter matrix refers to the camera projection parameter matrix including focal length and principal point offset. The first connection point refers to the precise spatial location of the end point of the blind zone entrance trajectory collected by the ground-based fixed monitoring equipment. This point serves as the starting anchor for the UAV to continue tracking from an aerial perspective, guiding the first UAV to the airspace where this point is located. Secondly, using the three-dimensional spatial coordinates of the first connection point as the center, a preset safe ground altitude is vertically offset directly above it. The three-dimensional position of the hovering point is then obtained. The attitude information of the hovering point and the gimbal pitch angle is locked at 90 degrees vertically downward as the hovering acquisition waypoint. The hovering acquisition waypoint refers to the information containing information used to guide the drone to hover. Then, when the police drone confirms that it has reached the radius threshold of the hovering acquisition waypoint, the ground control station triggers the airborne three-axis stabilization gimbal to rotate, so that the optical axis of the mounted visible light zoom camera is aligned with the vertical projection point of the three-dimensional spatial coordinates of the first connection point on the ground. The vertical projection point is used as the center of the image to continuously shoot the entrance area of ​​the monitoring blind spot, and obtain a multi-frame superimposed image of the blind spot entrance area. The blind spot entrance area image refers to the visible light digital image containing the surrounding environment of the location where the suspicious target disappears from the boundary of the monitoring blind spot. Finally, the blind spot entrance area image is input into the OSNet re-identification model pre-trained on the MSMT17 dataset. The re-identification model detects each pedestrian in the image and outputs the corresponding 512-dimensional re-identification feature vector. The feature vector with the highest cosine similarity to the received blind spot entrance feature is extracted as the first transition feature.

[0026] It should be noted that the first transitional feature in this application refers to the suspicious target re-identification feature vector recaptured from the perspective of the police drone. That is, the intermediate state feature parameter collected by the first drone at the first connection point through the airborne sensor and used to compare with the baseline feature of the blind zone entrance. The first transitional feature is used to characterize the instantaneous state of the suspicious target when it has just left the field of view of the ground fixed monitoring equipment and has not yet fully entered the depth of the monitoring blind zone. Its function is to serve as a feature connection between the ground monitoring perspective and the drone's aerial perspective, and to verify whether the drone has successfully connected to the correct target. In this way, the aerial-ground dual confirmation of the target identity is completed at the blind zone entrance, avoiding incorrect connection caused by sudden changes in perspective or target switching.

[0027] The guidance path determination module 200 is used to calculate the feature correlation degree between the first transition feature and the blind spot entrance feature. When the feature correlation degree is greater than a preset threshold, a blind spot crossing guidance path is determined based on the first connection point and the blind spot front trajectory, and the police drone is controlled to fly to the monitoring blind spot according to the blind spot crossing guidance path.

[0028] In some embodiments, reference Figure 3 As shown, this figure is an exemplary flowchart for determining a blind spot crossing guidance path according to this application. In this application, the first transition feature and the blind spot entrance feature are used to calculate the feature correlation degree. When the feature correlation degree is greater than a preset threshold, the blind spot crossing guidance path can be determined based on the first connection point and the trajectory before the blind spot using the following steps: In step 2001, the cosine similarity between the first transition feature and the blind zone entrance feature is calculated as the feature correlation degree; In step 2002, when the feature correlation degree is greater than a preset threshold, it is confirmed that the target corresponding to the first transition feature and the target corresponding to the blind zone entrance feature are the same suspicious target, and the motion direction vector corresponding to the blind zone pre-trajectory is determined; In step 2003, taking the first connection point as the starting point of path planning and the motion direction vector as the initial crossing direction, a smooth path curve is generated that passes through the monitoring blind zone and ends behind the monitoring blind zone as the blind zone crossing guidance path.

[0029] In specific implementation, firstly, the generated 512-dimensional first transition feature and the 2048-dimensional blind zone entry feature sent by the ground-based fixed monitoring equipment are simultaneously input into a domain adaptive feature mapping network consisting of three fully connected layers. The first two layers of this mapping network map the heterogeneous dimensional features to a unified 256-dimensional common subspace through shared weights, outputting the dimension-aligned first transition common feature and blind zone entry common feature. Then, the dot product between the two is calculated using the cosine similarity formula, divided by the modulus, to obtain a scalar value between -1 and 1 as the feature correlation degree. The domain adaptive feature mapping network refers to... A lightweight projection network, learned through joint training including source and target domain difference loss terms, is capable of eliminating the bias between airborne and fixed monitoring viewpoints. The feature correlation degree refers to the cross-domain identity consistency quantification value of the first transitional feature collected by the police drone and the blind spot entry feature sent by the ground-based fixed monitoring equipment. Then, when the feature correlation degree is greater than a preset threshold, it is confirmed that the airborne viewpoint target corresponding to the first transitional common feature and the fixed viewpoint target corresponding to the blind spot entry common feature are the same suspicious target. The preset threshold can be set to 0.75, which is not limited here. Subsequently, from the blind spot... The endpoint of the pre-blind zone trajectory is traced back along a predetermined range of trajectory points. The world coordinates of these traced trajectory points are then fitted with a least-squares three-dimensional straight line. The unit direction vector of the fitted straight line is calculated as the motion direction vector. This motion direction vector is a three-dimensional normalized pointing vector representing the tendency of the suspicious target to move in the final stage before entering the blind zone. Finally, the first connection point is used as the starting point for path planning. The angle between this motion direction vector and the ground plane is used as the initial pitch angle, and the projection of this motion direction vector onto the ground plane is used as the initial yaw angle. The coordinates of the prior exit area behind the blind zone are combined as the endpoint constraint. The quintic spline interpolation algorithm is invoked to generate a smooth path curve between the starting point and the ending point, with continuous curvature and a maximum curvature not exceeding the upper limit of the maneuverability of the police drone. This curve is the blind zone crossing guidance path. The quintic spline interpolation algorithm is a numerical method that uses the zeroth to second derivative values ​​at specified path points as boundary conditions to solve the polynomial coefficients of spline segments. The prior area coordinates of the exit refer to the three-dimensional world coordinates of the location of the most probable recurrence of the suspicious target in the area covered by the ground-based fixed monitoring equipment behind the monitoring blind zone, specifically based on historical trajectory statistics data. This will not be elaborated further here.

[0030] It should be noted that the blind spot crossing guidance path in this application refers to the three-dimensional spatial reference trajectory that guides the police drone to cross the monitoring blind spot along the most likely movement route of the target. Determining the blind spot crossing guidance path can provide the police drone with a three-dimensional spatial reference trajectory to cross the monitoring blind spot that its own sensors cannot directly cover. This allows the drone to use the most likely movement route of the suspected target as a priori guidance, and quickly cross the blind spot to reach the rear exit area while maintaining consistency with the predicted direction of the target's movement. This will compress the target loss time window to the shortest possible time and avoid the target re-acquisition delay caused by blind searching.

[0031] In the guidance path determination module, controlling the police drone to fly through the blind spot and reach the monitoring blind spot via the guidance path is achieved through the following steps: The blind spot traversal guidance path is discretized into a three-dimensional waypoint sequence with timestamps; The timestamped three-dimensional waypoint sequence is loaded into the flight path buffer of the police drone, triggering the police drone to perform a crossing maneuver in the waypoint sequence until it reaches the destination waypoint behind the monitoring blind spot.

[0032] In specific implementation, firstly, the smooth path curve represented by the blind zone crossing guidance path is parameterized and sampled with a fixed arc length step. For each sampling point, an arrival timestamp is assigned using a trapezoidal velocity planning algorithm based on the maximum cruising speed of the police drone and the path curvature radius, generating a three-dimensional waypoint sequence with timestamps containing spatial location and corresponding time. The trapezoidal velocity planning algorithm refers to a velocity allocation method that adaptively adjusts the desired speed according to the path curvature and back-calculates the cumulative time spent at each path point using a three-stage motion model of uniform acceleration, uniform speed, and uniform deceleration. This method is not limited here. The fixed arc length step can be set according to actual needs and is not limited here. Subsequently, the three-dimensional waypoint sequence with timestamps is written sequentially into the flight path buffer in the flash memory of the police drone's flight controller. Upon completion, the flight controller is triggered to switch to automatic route mode. The flight controller then drives the aircraft to perform crossing maneuvers in sequence according to the timestamp and three-dimensional coordinates of each waypoint through a cascaded PID position-velocity-attitude control loop. The route buffer refers to the structured circular queue storage area in the flight controller used to store preset route tasks. The crossing maneuver refers to the autonomous navigation action of the police drone flying along the waypoint connection line in sequence at a preset airspeed. When the flight controller determines that the sequence number of the currently reached waypoint is consistent with the sequence number of the destination waypoint and the horizontal position deviation is less than a preset threshold, it determines that the destination waypoint behind the monitoring blind zone has been reached. Then, it exits the automatic route mode and switches to a fixed-point hovering state to wait for the next stage instruction. The destination waypoint refers to the last three-dimensional waypoint marked as the end point of the path in the blind zone crossing guidance path.

[0033] The continuity feature recognition module 300 is used to acquire the blind zone exit feature behind the monitoring blind zone after the police drone flies to the monitoring blind zone, and to reverse associate the blind zone exit feature with the blind zone entrance feature and the first transition feature. If the reverse association is successful, the blind zone exit feature is marked as a continuity feature of the same suspicious target.

[0034] In the continuity feature recognition module, after the police drone flies to the monitoring blind spot, the acquisition of the blind spot exit feature behind the monitoring blind spot is achieved through the following steps: When the police drone reaches the end point of the guidance path through the blind spot, the onboard visual sensor of the police drone is controlled to perform a wide-area scan of the exit area behind the monitoring blind spot, centered on the three-dimensional spatial coordinates of the end point, and to obtain the video stream of the exit area. All moving targets appearing in the video stream of the export area are segmented into instances to generate a contour mask and its corresponding depth feature vector for each moving target. Calculate the feature correlation degree between each depth feature vector and the first transition feature, and filter out the moving target with the highest feature correlation degree as the suspected same suspicious target; Extract the pose key point heatmap, local texture gradient histogram and color space distribution vector of the suspected same suspicious target in the current frame, and then obtain the blind zone exit features behind the monitoring blind zone.

[0035] In specific implementation, firstly, after the police drone arrives at the endpoint of the guide path in the blind zone and hovers, the ground control station constructs a fan-shaped scanning area with the exit direction of the blind zone as the center line, using the three-dimensional spatial coordinates of the endpoint as the origin. By issuing gimbal commands, the airborne three-axis stabilization gimbal drives the visible light zoom camera to perform a line-by-line serpentine scan with a field of view of ±60 degrees horizontally and 0 to -45 degrees vertically, acquiring a continuous exit area video stream covering the entire exit area behind the monitoring blind zone. This exit area video stream refers to the visible light video frame sequence obtained by the police drone from the structured acquisition of the environment behind the blind zone at the endpoint. Secondly, the exit area video stream... The YOLOv8-seg instance segmentation model is input frame by frame. This model performs pixel-level classification and target detection for all pedestrians and vehicles in each frame, outputting a binary contour mask and corresponding bounding box for each moving target. The image region corresponding to each bounding box is then scaled to 256×256 pixels and input into the BoT re-identification backbone network. A 2048-dimensional depth feature vector is extracted from the global average pooling layer of this network. This generates a contour mask and its corresponding depth feature vector, each uniquely bound to a moving target. The contour mask is a single-channel binary image where target pixels are marked as 1 and background pixels as 0. The depth feature vector is obtained by... A compact numerical descriptor for identity comparison is extracted by a convolutional neural network. Then, the 512-dimensional first transition feature is input into a feature dimension mapping network consisting of two fully connected layers. The first layer of this network contains 1024 neurons and uses the ReLU activation function for non-linear transformation, while the second layer contains 2048 neurons and uses linear activation output, generating a 2048-dimensional mapped first transition feature. The feature dimension mapping network refers to a feature space projection structure pre-trained on the Market-1501 and MSMT17 joint datasets using cross-model feature alignment loss. The 2048-dimensional mapped first transition feature is then mapped to each moving target. The cosine similarity of each of the 2048-dimensional depth feature vectors of the target is calculated, and the moving target with the highest similarity value is taken as the suspected same suspicious target. Finally, the image regions of the suspected same suspicious target in the current frame are respectively fed into the ViTPose pose estimation model to extract the pose key point heatmap, into the texture analysis module based on the local binary mode operator to extract the local texture gradient histogram, and into the HSV color space conversion and histogram statistics algorithm to extract the color space distribution vector. Finally, these three heterogeneous feature vectors are concatenated along the channel dimension to generate a multimodal joint feature vector that integrates structural, texture and color multi-attribute information as the blind spot exit feature behind the monitoring blind spot.

[0036] It should be noted that, in this application, the blind zone exit feature refers to the appearance representation of the same suspicious target that may change due to changes in viewing angle, lighting differences, and partial occlusion after passing through the blind zone. Specifically, it is a multi-source descriptor that can be used for cross-domain identity comparison after jointly encoding the multi-dimensional appearance attributes of the suspected target in the exit area behind the blind zone. The blind zone exit feature refers to the multi-modal joint feature vector generated by fusing three heterogeneous appearance attributes—pose structure, texture details, and color distribution—for the suspected suspicious target recaptured by the police drone in the exit area behind the monitoring blind zone. This determines the blind zone exit feature, which can provide a richer and more complementary discrimination basis for subsequent cross-domain reverse association verification than a single deep feature. By locking the multi-dimensional appearance invariance of the target at the feature level, it is ensured that the target has not been replaced or lost only when the blind zone exit feature and the blind zone entrance feature and the first transition feature before entering the blind zone all constitute a high-confidence match. This solves the problem that relying solely on a single re-identification feature is unreliable for reverse association in complex outdoor scenarios, and provides a solid guarantee for the final labeling of the subsequent features.

[0037] In the continuity feature recognition module, the blind zone exit feature is back-associated with the blind zone entrance feature and the first transition feature. If the back-association is successful, the blind zone exit feature is marked as a continuity feature of the same suspicious target using the following steps: The blind zone exit feature and the blind zone entrance feature are input into the first twin network branch for cross-domain measurement to obtain the first reverse similarity score. At the same time, the blind zone exit feature and the first transition feature are input into the second twin network branch for cross-domain measurement to obtain the second reverse similarity score. The first reverse similarity score and the second reverse similarity score are weighted and fused to generate a reverse association confidence score; If the confidence level of the reverse association exceeds the preset confirmation threshold, the reverse association is determined to be successful, and the blind zone exit feature is marked as a successor feature of the same suspicious target.

[0038] In specific implementation, firstly, the blind zone exit feature is mapped to a 256-dimensional normalized exit projection feature through the first sub-coding network in the modal unified encoder; secondly, the blind zone entrance feature is mapped to a 256-dimensional normalized entrance projection feature through the second sub-coding network in the modal unified encoder; and thirdly, the first transition feature is mapped to a 256-dimensional normalized transition feature through the third sub-coding network in the modal unified encoder. The modal unified encoder refers to a heterogeneous feature composed of three structurally identical but parameter-independent four-layer fully connected sub-networks that compress original features of different dimensions and modalities into the same 256-dimensional common metric space. The alignment module is pre-trained on a joint dataset containing identity pairs labeled before and after the blind zone using triplet loss (details omitted here). The normalized exit projection features and the normalized entrance projection features are fed as input pairs into a first Siamese network branch. This first Siamese network calculates the Manhattan distance between the two input feature vectors and outputs a first inverse similarity score through a fully connected layer with a Sigmoid activation function. Simultaneously, the normalized exit projection features and the normalized transition features are fed as input pairs into a second Siamese network branch with identical structure and shared parameters. This second inverse similarity score is output using the same distance metric and mapping. Similarity scores, both the first and second reverse similarity scores, are scalar values ​​between 0 and 1, quantifying the identity consistency probability between exit and entry features, and between exit and transition features, in the common metric space, respectively. Then, the first and second reverse similarity scores are input into a weighted fusion module that determines the optimal weights through grid search. The reverse association confidence is calculated using a linear combination formula: the first reverse similarity score multiplied by 0.6 plus the second reverse similarity score multiplied by 0.4. The weighted fusion module refers to the calculation of the final [assumption / confidence] based on the entry and transition features on the historical validation set. The decision-level fusion unit uses a fixed weight coefficient preset based on the difference in contribution to identity discrimination. The reverse association confidence level refers to the overall confidence level of the consistency of the identity of the suspicious target obtained after integrating evidence from two independent association channels. If the reverse association confidence level exceeds the preset confirmation threshold, the reverse association is determined to be successful. The status flag bit in the memory structure corresponding to the blind zone exit feature is changed from pending confirmation to confirmed, so that it is marked as a successor feature of the same suspicious target. The confirmation threshold is a preset scalar critical value used to determine whether the reverse association is successful. It can be set according to actual needs or expert knowledge, and is not limited here.

[0039] It should be noted that, in this application, "continuing features" refers to legitimate appearance data of targets identified as belonging to the same suspicious target that reappear after crossing a monitoring blind zone. In existing technologies, UAV control and dispatch typically rely solely on appearance features from a single camera's perspective for recapture. Once a target enters a monitoring blind zone, abrupt changes in perspective, lighting, or partial occlusion can cause a break in appearance representation, easily leading to identity switching errors. In contrast, this solution's continuing features first confirm the target's identity from an aerial perspective to a fixed perspective through a positive correlation between the blind zone entrance feature and the first transition feature. Then, at the exit end, a multimodal joint feature integrating attitude structure, texture details, and color distribution is specifically generated as the blind zone exit feature. The blind zone exit feature is simultaneously back-linked to the blind zone entrance feature and the first transition feature, forming a closed-loop bidirectional verification from the exit back to the entrance and transition point. Only when the two independent twin network branches output high-confidence similarity scores and exceed the confirmation threshold after adaptive weighted fusion, is the blind zone exit feature marked as a successor feature of the same suspicious target. This makes the identification of successor features not a one-time feature distance comparison, but a decision-level confirmation result that dynamically evaluates the evidence weights of the two verification paths through an adaptive gating network. This effectively avoids target identity confusion caused by the randomness of single feature matching and achieves deterministic locking of the identity continuation of suspicious targets during the process of crossing the entire blind zone.

[0040] The flight path correction module 400 is used to perform spatiotemporal extrapolation of the movement trajectory of the suspicious target by means of the continuity features and the trajectory before the blind zone, thereby generating the complete trajectory of the suspicious target in the monitoring blind zone, and correcting the tracking and control flight path of the police drone based on the complete trajectory.

[0041] In the flight path correction module, the spatiotemporal extrapolation of the suspicious target's trajectory based on the continuity features and the trajectory before the blind zone, thereby generating the complete trajectory of the suspicious target within the monitoring blind zone, is achieved through the following steps: Extract the timestamps, three-dimensional spatial coordinates, and instantaneous velocity vectors of each trajectory point in the trajectory before entering the blind zone, and construct a temporal state sequence containing the change law of the motion state of the suspicious target before entering the monitoring blind zone; Using the time-series state sequence as input, the historical motion pattern of the suspected target is encoded using a long short-term memory network to generate motion pattern latent variables; Using the motion pattern latent variables as initial state constraints, and combined with the geographic information constraint layer within the monitoring blind zone, a continuous spatiotemporal pose sequence is generated step-by-step by a cyclic inference decoder from the end point timestamp of the trajectory before the blind zone to the timestamp when the connecting feature is first detected, as the complete trajectory of the suspicious target within the monitoring blind zone.

[0042] In specific implementation, firstly, the GPS timestamps of all trajectory points from the first point to the last point in the trajectory before the blind zone are taken, along with the three-dimensional spatial coordinates of the world coordinate system after total station calibration transformation, and the instantaneous velocity vectors of each point calculated by dividing the coordinate difference between adjacent trajectory points by the time difference. These three attributes of each trajectory point are concatenated into a nine-dimensional state vector, which is then arranged in ascending order by timestamp to construct a time-series state sequence containing the continuous change law of the movement state of the suspicious target before entering the monitoring blind zone. The time-series state sequence refers to a multivariate time series characterizing the evolution of the target's position and velocity over time. Then, the time-series state sequence is input into a large number of pedestrians. The encoder consists of a two-layer bidirectional long short-term memory network trained on vehicle trajectory data. Each layer of the encoder contains 128 hidden units. The first layer uses a forward LSTM to process the sequence in forward time order and a backward LSTM to process the sequence in reverse time order. The hidden states of the last time step of the forward and backward LSTM are concatenated into a 256-dimensional historical motion pattern encoding vector. The second layer uses this 256-dimensional vector as the initial hidden state to repeat the forward and backward encoding and output the final 256-dimensional motion pattern latent variable. The motion pattern latent variable refers to the latent space feature vector that compresses the speed change habits, steering preferences and gait rhythm of the suspicious target before entering the blind zone.Finally, the motion pattern latent variables are used as initial state constraints to input a recurrent inference decoder guided by an attention mechanism. This decoder consists of a single-layer LSTM unit, whose initial hidden state is set to the motion pattern latent variables and the initial cell state is set to zero. At each inference time step, the target 3D coordinates and velocity vectors predicted from the previous time step are concatenated and embedded through a fully connected layer as the input for the current step. Simultaneously, a geographic information constraint layer within the monitoring blind zone corrects the validity of the predicted output at each step. This geographic information constraint layer is composed of three types of data: the digital elevation model raster map corresponding to the monitoring blind zone, the building redline polygon vector map layer, and the topology of passable roads. For each candidate pose output by the decoder, the constraint layer performs point detection within the polygon to eliminate poses that penetrate into the building, performs an elevation nearest neighbor query to attach the pose to the ground, and performs road network distance calculation. Constraints are used to correct abnormal displacements deviating from the main road. Only the legal poses corrected by constraints are used as the formal output of that time step and fed into the next time step. The cyclic deduction decoder refers to a generative recursive network that generates future trajectory sequences step by step based on historical motion patterns and spatial environment constraints. The geographic information constraint layer refers to a rule and grid joint constraint module that performs physical rationality judgment and spatial projection correction on the generated poses in real time during trajectory deduction. Thus, starting from the end point timestamp of the trajectory before the blind zone, the deduction is performed step by step at equal time intervals until the timestamp when the connecting feature is first detected is reached. The three-dimensional coordinates and corresponding timestamps of all time steps generated in the entire process are collected and connected end to end to generate a spatiotemporal pose sequence that is continuous in both time and space and conforms to environmental physical constraints. This sequence is the complete trajectory of the suspicious target in the monitoring blind zone.

[0043] It should be noted that, in this application, the complete trajectory refers to a synthetic spatiotemporal path used to fill in the missing segments of the target's true trajectory within the monitoring blind zone. Determining the complete trajectory involves generating the most probable spatiotemporal path to fill the trajectory gaps in the blind zone through a deductive method that integrates behavioral pattern learning and geographic information constraints. This path serves as a bridge connecting the end point of the trajectory before the blind zone and the first detection point of the subsequent features after the blind zone, providing the police drone with a complete target motion prediction reference line that runs through the entire blind zone. This allows the drone to maintain a general prediction of the target's location based on this reference line even when the target is not visible, and to immediately switch to a corrected tracking route starting from the end of the complete trajectory when the target reappears in the exit area. This eliminates the interruption of drone trajectory planning and re-acquisition delay caused by the missing trajectory in the blind zone, ultimately achieving uninterrupted trajectory tracking and control of suspicious targets across the entire blind zone.

[0044] In the flight path correction module, the following steps are used to correct the tracking and control flight path of the police drone based on the completed trajectory: The completed trajectory is projected from the world coordinate system onto the onboard real-time local grid map of the police drone to generate a reference path layer containing all spatiotemporal poses of the suspicious target within the monitoring blind zone; Using the reference path layer as a planning constraint, a gradient-based trajectory optimizer generates a smooth tracking and control route that satisfies kinematic constraints and minimizes the average distance to the reference path layer between the current waypoint of the police drone and the latest detection position corresponding to the successive feature.

[0045] In specific implementation, firstly, the 3D world coordinates of all discrete spatiotemporal pose points in the completed trajectory are projected and transformed using a homogeneous transformation matrix from the world coordinate system to the airborne local grid map coordinate system. In the real-time local grid map maintained by the onboard computer of the police drone, each grid cell in which the projection point falls is assigned a reference path occupancy value. The set of all grid cells marked by the projection points together constitutes the reference path layer. The airborne real-time local grid map refers to the occupied grid map obtained by dividing a square area with a side length of 200 meters centered on the current hovering position of the police drone at a resolution of 0.5 meters. The reference path layer refers to the grid layer in the occupied grid map that identifies the most likely path traversed by a suspicious target with a specific value. Subsequently, this reference path layer is superimposed as a planning constraint layer onto the cost map of the airborne local trajectory planner. The trajectory planner constructs an initial B-spline curve between the current waypoint of the police drone and the latest detection position corresponding to the continuation feature. Using the control point coordinates of this B-spline curve as optimization variables, a gradient-based nonlinear trajectory optimizer is used for iterative optimization. Solution: The objective function of this optimizer consists of four weighted cost terms: the first term is the smoothing cost, which calculates the integral of the squares of accelerations between control points to penalize abrupt changes in path curvature; the second term is the kinematic constraint cost, which uses the maximum flight speed, maximum horizontal acceleration, and maximum yaw rate of the police drone as hard constraint boundaries, and uses the obstacle function method to pull the gradients of control point coordinates that exceed the constraints back into the feasible region; the third term is the obstacle collision cost, which is calculated by querying the grids occupied by buildings and trees in the local grid map and using the reciprocal of the distance from the control point to the nearest obstacle as the rejection factor. The potential gradient pushes the curve away from the obstacle; the fourth term is the path reference cost, which calculates the sum of the Euclidean distances from each sampling point of the B-spline curve to the center of the nearest reference path grid in the reference path layer. The gradient of this distance sum guides the optimizer to iterate in the direction of minimum distance. After each iteration, the control point coordinates are updated and the four costs are recalculated. After the cost decrease is lower than the convergence threshold, the optimal control point sequence is output. Then, the control point sequence is used to generate a tracking and control route with continuous curvature, satisfying all kinematic constraints and having the minimum average distance to the reference path layer through the B-spline interpolation formula.

[0046] It should be noted that the tracking and control route in this application refers to the guiding flight path used by police drones to track suspicious targets in blind spots. This allows the drone to actively advance along the path where the target is most likely to appear, thereby achieving close-range contact in spatial position the moment the target reappears from behind the blind spot. This eliminates the delay and gap in control caused by the route deviating from the actual movement path of the target, and completes the control closed loop from target loss to re-acquisition to continuous tracking.

[0047] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0048] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A grassroots police drone control and dispatch system, characterized in that, The system includes: The acquisition module is used to receive the trajectory of a suspicious target before entering the blind zone and the characteristics of the blind zone entrance sent by the ground fixed monitoring equipment, and to dispatch a police drone to collect the first transition features with the end point of the trajectory before entering the blind zone as the first connection point. The guidance path determination module is used to calculate the feature correlation degree between the first transition feature and the blind spot entrance feature. When the feature correlation degree is greater than a preset threshold, a blind spot crossing guidance path is determined based on the first connection point and the blind spot front trajectory, and the police drone is controlled to fly to the monitoring blind spot according to the blind spot crossing guidance path. The continuity feature recognition module is used to obtain the blind zone exit feature behind the monitoring blind zone after the police drone flies to the monitoring blind zone, and to reverse associate the blind zone exit feature with the blind zone entrance feature and the first transition feature. If the reverse association is successful, the blind zone exit feature is marked as the continuity feature of the same suspicious target. The flight path correction module is used to perform spatiotemporal extrapolation of the movement trajectory of the suspicious target by using the continuity features and the trajectory before the blind zone, thereby generating the complete trajectory of the suspicious target in the monitoring blind zone, and correcting the tracking and control flight path of the police drone based on the complete trajectory.

2. The grassroots police drone control and dispatch system as described in claim 1, characterized in that, In the acquisition module, the trajectory of a suspicious target before entering the blind zone and the characteristics of the blind zone entrance sent by the ground-based fixed monitoring equipment specifically include: Motion masks of suspicious targets are extracted from continuous video streams from ground-based fixed monitoring equipment. Target identifiers are assigned to the motion masks using a multi-target tracking algorithm, and trajectory segments with the target identifiers are generated. Determine whether the end point of the trajectory segment enters the preset monitoring blind zone boundary. When it enters the monitoring blind zone boundary, mark all trajectory points of the trajectory segment as the trajectory before the blind zone. Within a preset time window before the end point of the trajectory before the blind zone, depth features are extracted from the image block of the suspected target to generate blind zone entry features.

3. The grassroots police drone control and dispatch system as described in claim 1, characterized in that, In the data acquisition module, the process of dispatching the police drone to acquire the first transition feature using the end point of the trajectory before the blind spot as the first connection point specifically includes: Using the end point of the trajectory before the blind zone as the first continuation point, the first continuation point is converted from the pixel coordinate system of the ground fixed monitoring equipment to a unified world coordinate system, generating the three-dimensional spatial coordinates of the first continuation point. A hovering data acquisition waypoint is generated with the three-dimensional spatial coordinates of the first connection point as the center. After the police drone reaches the hovering data collection point, it uses the onboard visual sensor of the police drone to project images of the entrance area of ​​the monitoring blind spot in the direction corresponding to the three-dimensional spatial coordinates of the first connection point on the ground, thereby obtaining an image of the entrance area of ​​the blind spot. Suspicious targets in the blind zone entrance area image are identified by feature recognition to generate a first transition feature.

4. The grassroots police drone control and dispatch system as described in claim 1, characterized in that, In the guidance path determination module, the first transition feature and the blind spot entrance feature are used to calculate the feature correlation degree. When the feature correlation degree is greater than a preset threshold, the blind spot crossing guidance path is determined based on the first connection point and the blind spot pre-trajectory, specifically including: Calculate the cosine similarity between the first transition feature and the blind zone entrance feature as the feature correlation degree; When the correlation degree of the features is greater than a preset threshold, it is confirmed that the target corresponding to the first transition feature and the target corresponding to the blind zone entrance feature are the same suspicious target, and the motion direction vector corresponding to the trajectory before the blind zone is determined. Using the first connection point as the starting point of the path planning and the motion direction vector as the initial crossing direction, a smooth path curve is generated that passes through the monitoring blind zone and ends behind the monitoring blind zone as the blind zone crossing guidance path.

5. The grassroots police drone control and dispatch system as described in claim 1, characterized in that, In the guidance path determination module, controlling the police drone to fly through the blind spot and reach the monitoring blind spot specifically includes: The blind spot traversal guidance path is discretized into a three-dimensional waypoint sequence with timestamps; The timestamped three-dimensional waypoint sequence is loaded into the flight path buffer of the police drone, triggering the police drone to perform a crossing maneuver in the waypoint sequence until it reaches the destination waypoint behind the monitoring blind spot.

6. The grassroots police drone control and dispatch system as described in claim 1, characterized in that, In the continuity feature recognition module, after the police drone flies to the monitoring blind spot, acquiring the blind spot exit feature behind the monitoring blind spot specifically includes: When the police drone reaches the end point of the guidance path through the blind spot, the onboard visual sensor of the police drone is controlled to perform a wide-area scan of the exit area behind the monitoring blind spot, centered on the three-dimensional spatial coordinates of the end point, and to obtain the video stream of the exit area. All moving targets appearing in the video stream of the export area are segmented into instances to generate a contour mask and its corresponding depth feature vector for each moving target. Calculate the feature correlation degree between each depth feature vector and the first transition feature, and filter out the moving target with the highest feature correlation degree as the suspected same suspicious target; Extract the pose key point heatmap, local texture gradient histogram and color space distribution vector of the suspected same suspicious target in the current frame, and then obtain the blind zone exit features behind the monitoring blind zone.

7. The grassroots police drone control and dispatch system as described in claim 1, characterized in that, In the continuity feature recognition module, the blind zone exit feature is back-associated with the blind zone entrance feature and the first transition feature. If the back-association is successful, the blind zone exit feature is marked as a continuity feature of the same suspicious target. Specifically, this includes: The blind zone exit feature and the blind zone entrance feature are input into the first twin network branch for cross-domain measurement to obtain the first reverse similarity score. At the same time, the blind zone exit feature and the first transition feature are input into the second twin network branch for cross-domain measurement to obtain the second reverse similarity score. The first reverse similarity score and the second reverse similarity score are weighted and fused to generate a reverse association confidence score; If the confidence level of the reverse association exceeds the preset confirmation threshold, the reverse association is determined to be successful, and the blind zone exit feature is marked as a successor feature of the same suspicious target.

8. The grassroots police drone control and dispatch system as described in claim 1, characterized in that, In the flight path correction module, correcting the tracking and control flight path of the police drone based on the completed trajectory specifically includes: The completed trajectory is projected from the world coordinate system onto the onboard real-time local grid map of the police drone to generate a reference path layer containing all spatiotemporal poses of the suspicious target within the monitoring blind zone; Using the reference path layer as a planning constraint, a gradient-based trajectory optimizer generates a smooth tracking and control route that satisfies kinematic constraints and minimizes the average distance to the reference path layer between the current waypoint of the police drone and the latest detection position corresponding to the successive feature.

Images