Air-water amphibious unmanned aerial vehicle automatic switching search and rescue system and method

Abstract

The invention relates to an air-water amphibious unmanned aerial vehicle automatic switching search and rescue system and method, and belongs to the technical field of overwater emergency rescue. The air-water amphibious unmanned aerial vehicle automatic switching search and rescue system comprises an unmanned aerial vehicle automatic switching search and rescue module and an intelligent emergency control module; the unmanned aerial vehicle automatic switching search and rescue module is used for collecting image data and point cloud data of a search and rescue area in real time according to a search and rescue task and transmitting the image data and the point cloud data to the intelligent emergency control module; and the intelligent emergency control module is used for identifying a search and rescue target based on the image data and the point cloud data, tracking the identified search and rescue target, obtaining a target pose, planning a flight path of the unmanned aerial vehicle self-switching search and rescue module based on the target pose and the obtained pose of the unmanned aerial vehicle self-switching search and rescue module, and sending the planned flight path to the unmanned aerial vehicle self-switching search and rescue module. And the unmanned aerial vehicle automatic switching search and rescue module is controlled to land around the search and rescue target based on the planned flight path, so that air-water amphibious unmanned aerial vehicle automatic switching search and rescue are realized, and the timeliness and reliability of emergency rescue are improved.

Description

Technical Field

[0001] This invention relates to the field of water emergency rescue technology, and in particular to an air-water amphibious unmanned aerial vehicle (UAV) self-deployment search and rescue system and method. Background Technology

[0002] With the increasing frequency of inland waterway and maritime traffic, the risk of maritime safety accidents continues to rise. Traditional search and rescue methods mainly rely on rescue vessels and human visual inspection, which have problems such as long response delays, limited search range, difficulty in approaching targets in bad sea conditions, and a sharp drop in efficiency at night or in low visibility environments.

[0003] Existing UAV visual recognition and detection methods are susceptible to interference in complex water environments. From the high-altitude aerial perspective of UAVs, the target pixel ratio of a person in the water is small. Furthermore, under extreme conditions such as high waves and strong light, as well as interference from sea surface ripples and floating objects, severe background interference is easily generated. Strong reflections and ripples on the water surface can easily form highlight areas, severely distorting image quality. The person in the water may be submerged by reflections or confused with floating debris, obscuring human features and making it difficult to accurately identify the person in distress and obstacles on the water surface. Moreover, after the UAV identifies a person in the water, it transmits the information to the rescue vessel, preventing the person from quickly escaping danger and reducing the timeliness and reliability of maritime emergency rescue. Summary of the Invention

[0004] In view of this, it is necessary to provide an air-water amphibious unmanned aerial vehicle (UAV) self-deployment search and rescue system and method to solve the technical problems of insufficient target detection accuracy and low timeliness and reliability of emergency rescue in complex water environments.

[0005] To address the aforementioned problems, in a first aspect, the present invention provides an air-water amphibious unmanned aerial vehicle (UAV) self-deployment search and rescue system, comprising an UAV self-deployment search and rescue module and an intelligent emergency control module; The drone self-deployment search and rescue module is used to collect image data and point cloud data of the search and rescue area in real time according to the search and rescue mission, and transmit the image data and point cloud data to the intelligent emergency control module. The intelligent emergency control module is used to identify the search and rescue target based on the image data and point cloud data, track the identified search and rescue target, obtain the target pose, plan the flight trajectory of the UAV self-deployed search and rescue module based on the target pose and the obtained pose of the UAV self-deployed search and rescue module, and control the UAV self-deployed search and rescue module to land near the search and rescue target based on the planned flight trajectory, so as to realize the self-deployed search and rescue of the air and water amphibious UAV.

[0006] In one possible implementation, the UAV self-deployed search and rescue module includes a multimodal camera unit and a communication unit; The multimodal camera unit is used to collect image data and point cloud data of the search and rescue area in real time according to the search and rescue mission; The communication unit is used to transmit the image data and point cloud data to the intelligent emergency control module.

[0007] In one possible implementation, the UAV self-deployed search and rescue module further includes a composite lightweight airframe structure, a hybrid propulsion unit, an energy and management unit, and a propeller protection net. The composite lightweight body structure is in the shape of a concentric circle lifebuoy. The edge of the concentric circle lifebuoy includes an ergonomic concave surface for supporting the search and rescue target and a ring-shaped functional handle for the search and rescue target to grip. The hybrid propulsion unit is used to switch between high-speed aerial cruising and agile maneuvering on the water. The energy and management unit is used to provide energy supply for the multimodal camera unit, communication unit, and hybrid propulsion unit; The propeller protection net is used to provide safety protection for the drone's propellers.

[0008] In one possible implementation, the intelligent emergency control module includes an intelligent analysis unit, a trajectory planning unit, and a control unit; The intelligent analysis unit is used to perform search and rescue target detection on the image data using a deep learning model, obtain the category, detection box, and image texture features of the search and rescue target, extract the point cloud geometric features of the search and rescue target based on the point cloud data, match the point cloud geometric features with the image texture features to determine the search and rescue target and the search and rescue target pose, and use Kalman filtering to track the search and rescue target to obtain the target pose. The network architecture of the deep learning model is the YOLOv11n model. The trajectory planning unit is used to plan the flight trajectory of the UAV self-deployment search and rescue module based on the target pose and the acquired pose of the UAV self-deployment search and rescue module. The control unit is used to control the UAV self-deployment search and rescue module to land near the search and rescue target based on the planned flight trajectory, so as to realize the self-deployment search and rescue of the air and water amphibious UAV.

[0009] In one possible implementation, the YOLOv11n model includes a backbone network, a neck network, a DyHead layer, and a detection head, and the intelligent analysis unit is specifically used for: The image data is subjected to multi-scale feature extraction through the backbone network to obtain first-scale features, second-scale features and third-scale features. The first-scale feature, the second-scale feature, and the third-scale feature are fused using the neck network to obtain fused image features; The fused image features are enhanced through the attention mechanism of the DyHead layer, wherein the attention mechanism includes a scale-aware attention mechanism, a spatial-aware attention mechanism, and a task-aware attention mechanism. The enhanced image features are analyzed by a detection head to perform category prediction, bounding box regression, and target prediction, thereby obtaining the target's category, detection box, and image texture features.

[0010] In one possible implementation, the loss function of the deep learning model is: , , , , , , , in, The loss function of the deep learning model. The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. For distance loss, For angle loss, For the channel height, For channel width, For the first One detection box, For the first One detection box, For the predicted bounding box and the ground truth bounding box in Normalized distance in the direction, For the predicted bounding box and the ground truth bounding box in Normalized distance in the direction, The correlation coefficient for angle loss, The correlation coefficient for angle loss, This is for shape loss.

[0011] In one possible implementation, the intelligent analysis unit is further used for: Calculate the cosine similarity between the point cloud geometric features and the image texture features; When the cosine similarity is greater than or equal to a preset similarity threshold, a correspondence between point cloud geometric features and image texture features is established. The point cloud geometric features corresponding to the image texture features are fused to determine the search and rescue target and its pose.

[0012] In one possible implementation, the trajectory planning unit is specifically used for: The starting point and ending point of the UAV self-deployed search and rescue module are determined based on the target pose and the pose of the UAV self-deployed search and rescue module. Based on the starting point and the ending point, a cubic B-spline smoothing curve is used to plan the flight trajectory of the UAV self-deployment search and rescue module, and the planned flight trajectory is obtained.

[0013] In one possible implementation, the intelligent emergency control module further includes a communication and navigation unit, a wireless charging unit, a multi-dimensional monitoring unit, and a charging pile interface unit; The communication and navigation unit is used to perform real-time data interaction and centimeter-level positioning with the UAV self-deployment search and rescue module. The wireless charging unit is used to provide wireless charging for the UAV self-deployment search and rescue module. The multi-dimensional monitoring unit is used to monitor the temperature, humidity, and wind speed of the wireless charging unit; The charging pile interface unit is used to provide an installation interface for the communication and navigation unit, the wireless charging unit, and the multi-dimensional monitoring unit.

[0014] Secondly, the present invention also provides a self-deployed search and rescue method for amphibious unmanned aerial vehicles (UAVs), implemented through the self-deployed search and rescue system for amphibious UAVs described above, comprising: Based on the UAV self-deployment search and rescue module, image data and point cloud data of the search and rescue area are collected in real time according to the search and rescue mission, and the image data and point cloud data are transmitted to the intelligent emergency control module. The intelligent emergency control module identifies the search and rescue target based on the image data and point cloud data, tracks the identified target, obtains the target pose, plans the flight trajectory of the UAV self-deployed search and rescue module based on the target pose and the obtained pose of the UAV self-deployed search and rescue module, and controls the UAV self-deployed search and rescue module to land near the search and rescue target based on the planned flight trajectory, realizing the self-deployed search and rescue of the air and water amphibious UAV.

[0015] The beneficial effects of this invention are as follows: The air-sea amphibious UAV self-drop search and rescue system includes a UAV self-drop search and rescue module and an intelligent emergency control module. Based on the search and rescue mission, it collects image data and point cloud data of the search and rescue area in real time and transmits the image data and point cloud data to the intelligent emergency control module. The intelligent emergency control module identifies the search and rescue target based on the image data and point cloud data, tracks the identified target, obtains the target's pose, and plans the flight trajectory of the UAV self-drop search and rescue module based on the target's pose and the acquired pose of the UAV self-drop search and rescue module. Based on the planned flight trajectory, it controls the UAV self-drop search and rescue module to land near the search and rescue target, realizing air-sea amphibious UAV self-drop search and rescue. By identifying the search and rescue target through image data and point cloud data, it effectively eliminates interference from sea surface waves and floating objects, accurately identifies the search and rescue target and water surface obstacles, improves target detection accuracy, and controls the UAV to land near the target according to the planned trajectory, helping people who have fallen into the water to quickly escape danger, improving the timeliness and reliability of emergency rescue. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A schematic diagram of an embodiment of the self-deployed search and rescue system for an air-water amphibious unmanned aerial vehicle provided by the present invention; Figure 2 This is a schematic diagram of the structure of a drone self-deployment search and rescue module of the air-water amphibious drone self-deployment search and rescue system provided by the present invention. Figure 3 A schematic diagram of the structure of an embodiment of the intelligent emergency control module of the air-water amphibious unmanned aerial vehicle self-deployment search and rescue system provided by the present invention; Figure 4 A schematic flowchart of an embodiment of the self-deployed search and rescue method for an air-water amphibious unmanned aerial vehicle provided by the present invention; Figure 5 A schematic diagram of the overall structure of the air-water amphibious UAV self-deployment search and rescue method provided by the present invention. Figure 6 A schematic diagram of the structure of the self-deployed search and rescue UAV in the air-water amphibious UAV self-deployed search and rescue method provided by the present invention; Figure 7 This is a schematic diagram of the intelligent emergency control and communication center for the self-deployed search and rescue method of the air-water amphibious UAV provided by the present invention. Detailed Implementation

[0018] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] This invention discloses a self-deployed search and rescue system for an amphibious unmanned aerial vehicle (UAV). Figure 1 This is a schematic diagram of the self-deployed search and rescue system for an amphibious unmanned aerial vehicle (UAV) provided in an embodiment of the present invention. Please refer to [link / reference]. Figure 1 The air-water amphibious unmanned aerial vehicle (UAV) self-deployment search and rescue system includes an UAV self-deployment search and rescue module 1100 and an intelligent emergency control module 1200. The drone self-deployment search and rescue module 1100 is used to collect image data and point cloud data of the search and rescue area in real time according to the search and rescue mission, and transmit the image data and point cloud data to the intelligent emergency control module 1200. The intelligent emergency control module 1200 is used to identify search and rescue targets based on image data and point cloud data, track the identified search and rescue targets, obtain the target pose, plan the flight trajectory of the UAV self-deployed search and rescue module 1100 based on the target pose and the obtained pose of the UAV self-deployed search and rescue module 1100, and control the UAV self-deployed search and rescue module 1100 to land near the search and rescue target based on the planned flight trajectory, so as to realize the self-deployed search and rescue of the air and water amphibious UAV.

[0021] By combining the drone self-deployment search and rescue module and the intelligent emergency control module, the accuracy of target detection is improved, helping people who have fallen into the water to quickly escape danger and improving the timeliness and reliability of emergency rescue.

[0022] In some embodiments, please refer to Figure 2 The drone self-deployed search and rescue module 1100 includes a multi-modal camera unit 1110 and a communication unit 1120; The multimodal camera unit 1110 is used to collect image data and point cloud data of the search and rescue area in real time according to the search and rescue mission; The multimodal camera unit 1110 includes a visible light high-definition camera and an infrared camera radar. It simultaneously records multi-view images using the visible light high-definition camera and the infrared camera radar. The overlap of the built-in dual-camera fields of view of the multimodal camera must meet the following requirements: , in, This represents the area of ​​overlap between the fields of view of the visible light high-definition camera and the infrared camera radar. This is the sum of the field of view areas of the visible light high-definition camera and the infrared camera radar. The preset field-of-view overlap threshold is typically set above 0.75 to ensure information complementarity between the two perception modes. The infrared camera radar uses the phase difference method to measure target distance. , in, For the target distance, At the speed of light, To achieve laser modulation frequency, the high-definition camera and the infrared camera radar are synchronized and triggered for data acquisition via a hardware clock, ensuring that the time error meets the specified requirements. , The maximum permissible synchronization error is determined by the sensor's response speed and is typically no more than 10 milliseconds. Communication unit 1120 is used to transmit image data and point cloud data to the intelligent emergency control module; The UAV self-deployed search and rescue module 1100 also includes a composite lightweight airframe structure 1130, a hybrid propulsion unit 1140, an energy and management unit 1150, and a propeller protection net 1160. The composite lightweight body structure 1130 is in the shape of a concentric circle lifebuoy. The edge of the concentric circle lifebuoy includes an ergonomic concave surface for supporting the search and rescue target and a ring-shaped functional handle for the search and rescue target to grip. The composite lightweight airframe structure is watertight and corrosion-resistant, and has undergone fluid simulation optimization to reduce wind and water resistance. It has a large surface area and buoyancy, ensuring that the drone can be moored on the water surface and meeting the needs of rescuing people who have fallen into the water. The hybrid propulsion unit 1140 is used to switch between high-speed aerial cruising and agile maneuvering on the water surface of the UAV. The hybrid propulsion unit 1140 includes an aerial propulsion unit and a water propulsion unit. The aerial propulsion unit uses a dual-redundant brushless motor and a propeller with optimized aerodynamic design, while the water propulsion unit uses an omnidirectional vectoring jet pump or a servo-controlled thruster, supporting high-speed aerial cruising and water surface operation. It is flexible and maneuverable, enabling seamless switching between high-speed aerial cruising and flexible maneuverability on the water. The energy and management unit 1150 is used to provide energy supply for the multimodal camera unit 1110, the communication unit 1120, and the hybrid propulsion unit 1140; The Energy and Management Unit 1150 uses a 4100mAh triple-cell LiPo battery pack as the main power source and integrates a wireless charging module as an auxiliary power source. It has built-in modules including current limiting protection, soft start, power monitoring and multiple step-down circuits to ensure stable, safe and long-lasting operation of the system. The propeller protection net 1160 is used to provide safety protection for the propellers of drones; The 1160 propeller protection net is made of high-strength, lightweight alloy material in a single piece, resulting in a robust yet lightweight structure that does not excessively increase the drone's load. Its carefully designed hexagonal honeycomb mesh layout ensures effective protection while minimizing its impact on aerodynamic performance. The connection points with the drone fuselage utilize an elastic buffer structure to effectively absorb the impact of external collisions. The surface also features a special coating that provides excellent corrosion resistance and wear resistance, allowing for long-term use in complex marine or aquatic environments. This provides reliable protection for the propellers and ensures drone flight safety.

[0023] In some embodiments, please refer to Figure 3 The intelligent emergency control module 1200 includes an intelligent analysis unit 1210, a trajectory planning unit 1220, and a control unit 1230; The intelligent analysis unit 1210 is used to perform search and rescue target detection on image data using a deep learning model, obtain the category, detection box, and image texture features of the search and rescue target, extract the point cloud geometric features of the search and rescue target based on point cloud data, match the point cloud geometric features with the image texture features to determine the search and rescue target and the search and rescue target pose, and use Kalman filtering to track the search and rescue target and obtain the target pose. The network architecture of the deep learning model is the YOLOv11n model, which includes a backbone network, a neck network, a DyHead layer, and a detection head. The image data is processed through a backbone network to extract multi-scale features, obtaining first-scale, second-scale, and third-scale features. The neck network then fuses these features to obtain fused image features. The fused image features are further enhanced using attention mechanisms in the DyHead layer, including scale-aware, spatial-aware, and task-aware attention mechanisms. Finally, a detection head performs category prediction, bounding box regression, and target prediction on the enhanced image features to obtain the target's category, detection box, and image texture features. Calculate the cosine similarity between point cloud geometric features and image texture features; when the cosine similarity is greater than or equal to a preset similarity threshold, establish a correspondence between point cloud geometric features and image texture features; fuse the corresponding point cloud geometric features and image texture features to determine the search and rescue target and its pose. The trajectory planning unit 1220 is used to plan the flight trajectory of the UAV self-deployment search and rescue module based on the target pose and the acquired pose of the UAV self-deployment search and rescue module. The starting point and ending point of the UAV self-deployed search and rescue module are determined based on the target pose and the pose of the UAV self-deployed search and rescue module. Based on the starting point and ending point, the flight trajectory of the UAV self-deployed search and rescue module is planned using a cubic B-spline smooth curve to obtain the planned flight trajectory. The control unit 1230 is used to control the UAV self-deployed search and rescue module to land near the search and rescue target based on the planned flight trajectory, so as to realize the self-deployed search and rescue of the air and water amphibious UAV.

[0024] The intelligent emergency control module 1200 also includes a communication and navigation unit 1240, a wireless charging unit 1250, a multi-dimensional monitoring unit 1260, and a charging pile interface unit 1270; The communication and navigation unit 1240 is used for real-time data interaction and centimeter-level positioning with the UAV self-deployment search and rescue module 1100; The communication and navigation unit 1240 integrates GPS / RTK high-precision positioning, data transmission radio, and 4G / 5G communication link to achieve real-time data interaction and centimeter-level positioning with the drone, and supports 4G / 5G, Wi-Fi and Bluetooth communication methods. The wireless charging unit 1250 is used to provide wireless charging for the drone self-deployed search and rescue module 1100; The multi-dimensional monitoring unit 1260 is used to monitor the temperature, humidity and wind speed of the wireless charging unit 1250; The charging pile interface unit 1270 provides an installation interface for the communication and navigation unit 1240, the wireless charging unit 1250, and the multi-dimensional monitoring unit 1260.

[0025] In some embodiments, the present invention discloses a self-deployed search and rescue method for an amphibious unmanned aerial vehicle (UAV). Figure 4 This is a flowchart of the self-deployed search and rescue method for an air-water amphibious unmanned aerial vehicle provided in an embodiment of the present invention, as follows: Figure 4 As shown, the self-deployed search and rescue method for amphibious drones includes: S401. The drone self-deployment search and rescue module collects image data and point cloud data of the search and rescue area in real time according to the search and rescue mission, and transmits the image data and point cloud data to the intelligent emergency control module. It should be noted that the multimodal camera unit of the drone's self-deployed search and rescue module simultaneously collects image data and point cloud data, providing a basis for identifying the target of the person who has fallen into the water.

[0026] S402: The intelligent emergency control module identifies the search and rescue target based on image data and point cloud data, tracks the identified search and rescue target, obtains the target pose, plans the flight trajectory of the UAV self-deployed search and rescue module based on the target pose and the obtained pose of the UAV self-deployed search and rescue module, and controls the UAV self-deployed search and rescue module to land around the search and rescue target based on the planned flight trajectory, so as to realize the self-deployed search and rescue of the air and water amphibious UAV. It should be noted that by extracting the category, detection box, and image texture features of the target of the person who fell into the water through a deep learning model, and by matching the point cloud with the target image, the interference of sea waves and floating objects is effectively eliminated, and the person who fell into the water and obstacles on the water surface are accurately identified, thus improving the target detection accuracy. According to the planned trajectory, the air-sea amphibious UAV is controlled to land around the search and rescue target, helping the person who fell into the water to get out of danger quickly, thus improving the timeliness and reliability of emergency rescue.

[0027] In some embodiments, in step S401, image data and point cloud data of the search and rescue area are collected in real time according to the search and rescue mission. The drone's self-deployed search and rescue module is an amphibious drone. After receiving the search and rescue mission, the amphibious drone quickly arrives at the search and rescue area and divides it into multiple regular sub-regions. The drone fully considers its flight speed, endurance, and signal transmission distance to ensure that each sub-region is within its effective operating range. After completing the area division, the amphibious drone searches each sub-region in a sequence from the edge to the center, or from the center outwards. During flight, the drone... The amphibious drone's multimodal camera scans the area below continuously and from all directions. After detecting a suspected target using a breadth-first search mode, it immediately switches to a precision search mode. The high-resolution optical camera captures high-definition images of the suspected target area, collecting image data. The infrared camera radar emits a laser beam, and by measuring the time it takes for the laser to reflect back, it accurately maps the three-dimensional spatial information around the target, thereby obtaining point cloud data. This allows it to determine parameters such as the target's position, altitude, and distance from the drone. After obtaining the image data and point cloud data, it transmits them to the intelligent emergency control module.

[0028] In some embodiments, in step S402, the search and rescue target is identified based on image data and point cloud data. A deep learning model is used to detect the target in the image data to obtain the category, detection box, and image texture features of the search and rescue target. The network architecture of the deep learning model is the YOLOv11n model, which includes a backbone network, a neck network, a DyHead layer, and a detection head. The backbone network extracts multi-scale features from the image data to obtain first-scale features, second-scale features, and third-scale features. The output of the backbone network is a three-dimensional tensor. The features are dimensional (L), spatial (S), and channel (C). The first-scale, second-scale, and third-scale features are fused through a neck network to obtain fused image features. The fused image features are then enhanced through the attention mechanism of the DyHead layer. This attention mechanism includes scale-aware, spatial-aware, and task-aware attention mechanisms. DyHead integrates these attention mechanisms, and by introducing attention mechanisms into each dimension of the feature tensor, it can unify and enhance feature representation capabilities. The attention function for a three-dimensional feature tensor is: , in, , , Attention mechanisms are applied in dimensions L, S, and C, respectively. These attention mechanisms are sequentially applied to the detection head and can be stacked multiple times to further enhance feature representation capabilities, improve computational efficiency, and enhance the real-time target detection capability for small targets on UAVs. The detection head performs category prediction, bounding box regression, and target prediction on the enhanced image features to obtain the target's category, detection box, and image texture features. The loss function of the deep learning model is: , , , , , , , in, The loss function of the deep learning model. The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. For distance loss, For angle loss, For the channel height, For channel width, For the first One detection box, For the first One detection box, For the predicted bounding box and the ground truth bounding box in Normalized distance in the direction, For the predicted bounding box and the ground truth bounding box in Normalized distance in the direction, The correlation coefficient for angle loss, The correlation coefficient for angle loss, For shape loss, when By retaining the detection boxes with the highest confidence, a one-to-one correspondence between the real target and the detection result can be achieved, effectively eliminating redundant duplicate detections and making the detection output more consistent with the target distribution characteristics of the actual scene. This screening mechanism provides high-precision data support for the stable operation of subsequent target tracking algorithms and the scientific formulation of rescue decisions. Compared with the processing scheme of directly retaining all detection boxes, the non-maximum suppression algorithm, based on threshold constraint screening logic, significantly reduces the sample size of subsequent data processing and significantly reduces the overall computational complexity. Especially in resource-constrained embedded platforms such as single UAVs, the elimination of redundant detection boxes can effectively reduce the processor's computational load and storage overhead, reserving sufficient hardware resources for the efficient execution of other core tasks such as navigation and communication, and ensuring the real-time performance and reliability of the overall system operation. Before using a deep learning model for object detection in image data, a structured pruning operation is performed on the deep learning model based on the constructed dependency graph and a preset pruning rate. In the detection of drowning victims, the complexity and number of parameters of the deep learning model continue to increase, placing a huge burden on the model's deployment and inference efficiency. To optimize the detection algorithm architecture and make the algorithm model lightweight, an empty dependency graph matrix is ​​first constructed. The dependency graph matrix is ​​used to store the dependency relationships between the layers of the network. Then, starting from the initial layer of the network, the inter-layer connection logic is analyzed layer by layer. For each layer, its input data and output results are checked to see if there are any related links with the inputs and outputs of other layers. If a valid connection exists, the corresponding position in the dependency graph matrix is ​​assigned a value of 1. Finally, a depth-first search algorithm is used, starting from the specified starting node and following the path formed by the dependency relationships to continuously traverse until all related nodes have been visited. The dependency graph is as follows: , in, For feature maps, and For inter-layer dependency, and This involves defining intra-layer dependencies. The inter-layer and intra-layer dependencies are used to determine the mutual dependencies between the inputs and outputs of each layer. Model compression is achieved by sparsifying the model parameters. The parameters are grouped according to the inter-layer dependencies to form a dependency graph. Then, group-level sparsity is achieved by introducing a regularization term into the loss function. The loss function for sparse training is: , in, The loss function of the original model. The sparsity factor is used to control the degree of sparsity. This is a regularization term used to measure the parameters. The importance of regularization terms is as follows: , in, For parameters The One predictable dimension for The square of the L2 norm, which is the sum of the squares of all its elements. Shrinkage strength is used to adjust the degree of sparsity in different dimensions. The shrinkage strength is: , in, and For parameter group The maximum and minimum importance scores, This is a hyperparameter used to control the range of shrinkage intensity. During sparse training, non-critical parameters in the model are gradually set to zero, thus forming a network structure with sparse characteristics. After sparse training, combined with the preset pruning ratio and dependency graph, parameter groups with zero channel weights can be selectively removed to complete the structured pruning operation of the model. After sparse training, channels with zero weights in the model will be clearly identified. Combined with the preset pruning rate and the constructed dependency graph, zero-weight channels and their associated upstream and downstream links can be pruned to complete the pruning operation. This can effectively reduce the number of parameters while ensuring that the original accuracy of the model remains basically unchanged, allowing the model to be deployed in a lightweight manner.

[0029] Considering the complex scenario of multiple people falling into the water, an improved density clustering algorithm is used to optimize the anchor frames. Based on the labeled data, the target size distribution is statistically analyzed, and the cluster centers are calculated. The anchor frame size is adjusted by measuring distance to improve its adaptability to people of different body types who fall into the water. The distance measurement is as follows: , in, , For the width and height of the target bounding box, , To determine the width and height of the cluster center boxes, the width and height information of the drowning group is extracted from the image taken by the drone using detection boxes. After standardization, the rationality analysis of the clustering results is performed to determine the most suitable number and size of detection boxes, which significantly improves the model's adaptability and detection accuracy to differences in the size and width-to-height ratio of the drowning group.

[0030] Geometric features of the search and rescue target are extracted from point cloud data. These geometric features are then matched with image texture features to determine the target and its pose. The cosine similarity between the point cloud geometric features and the image texture features is calculated. The cosine similarity is as follows: , in, For image texture feature vectors, The geometric feature vectors of the point cloud. For matching degree; when the cosine similarity is greater than or equal to the preset similarity threshold, the correspondence between point cloud geometric features and image texture features is established, and the corresponding point cloud geometric features and image texture features are fused to determine the target and target pose. By fusing the three-dimensional geometric contour of the point cloud with the texture features of the image, a target model with both spatial accuracy and visual recognition is constructed to determine the search and rescue target and its pose, providing accurate basic data for subsequent target tracking, search and rescue path planning and other tasks. This enables search and rescue drones to stably identify and locate search and rescue targets (people who have fallen into the water), obstacles and other targets in complex water environments, improving emergency response efficiency and search and rescue success rate.

[0031] Kalman filtering is used to track search and rescue targets and obtain their poses. The Kalman filtering algorithm is then used to achieve target tracking through state prediction and update equations.

[0032] The attitude of the air-sea amphibious UAV is obtained. A GPS and RTK fusion positioning method is used to determine the UAV's position. Combining sensor characteristics, information from the accelerometer, gyroscope, and magnetometer is used to update the UAV's attitude, position, and flight speed in real time using Kalman filtering. The Kalman filter state equation is: , in, It is a 7-dimensional state vector. , , , , For attitude quaternions, , , For gyroscope drift; the discrete model is: , , in, , For system process noise, For the attitude state transition matrix, the quaternion dynamic equations use the Picard first-order approximation equation. Let the solution period be... Then the attitude state transition matrix is: , in Because of the high update frequency, the gyroscope deviation can be assumed to be the same as the previous moment. Taking the predicted value from the previous time step, the state transition matrix is: , Covariance prediction is: , in, The covariance of the system process noise; after establishing the state equation and measurement equation, the equation is updated by Kalman filtering, with the gain being: , in, To measure the noise covariance, the state estimate is: , The covariance is estimated as follows: , After obtaining the UAV's attitude information, the transformation matrix from the body coordinate system to the geographic coordinate system can be used. Obtain the new body coordinate system to geographic coordinate system transformation matrix This allows for speed and position updates of the drone, with the speed update being... , Location updated to: , The attitude, position, and flight speed of the air-sea amphibious UAV are updated in real time by Kalman filtering, and the attitude of the air-sea amphibious UAV is obtained in real time. The flight trajectory of the amphibious drone is planned based on the real-time target pose and the real-time pose of the drone. The start and end points of the drone are determined based on these real-time poses, with the location of the target in the water as the end point and the drone's location as the start point. Based on the start and end points, a cubic B-spline smoothed curve is used to plan the drone's flight trajectory. Environmental modeling is performed considering factors such as the original terrain and obstacle areas. The baseline terrain model is as follows: , in, and Let these be the coordinates of the points projected onto the horizontal plane by the model. This represents the elevation value corresponding to a point on the horizontal plane. , , , , , , A constant coefficient controls the baseline terrain undulation in the digital map; the UAV's flight path is represented by ordered point coordinates, and the spatial coordinates that are close to each other are associated with each other using a cubic B-spline smoothing curve, assuming a set of node sequences. Describe each intermediate node; this sequence of nodes consists of... It consists of nodes. and These represent the start and end points of the drone, respectively. The three-dimensional representation of the various nodes in the flight process, including the takeoff point and the termination point, is as follows: , , The intermediate nodes are: , The purpose of the constraints is to ensure a flyable path is planned. Therefore, two constraints are used: terrain and environment. To avoid collisions during the information acquisition task, the UAV's flight altitude should always be higher than the terrain altitude. Therefore, the terrain constraint is modeled as follows: , in, This is a terrain function used to return the location. The terrain elevation value is used to determine the environmental constraints for the drone during information acquisition tasks. To plan a better path and reduce costs, the drone is restricted to operating only within a designated area. The environmental constraint model is as follows: , For flyable paths, considering the UAV's range, obstacles, and boundary constraints, the comprehensive cost function of the aircraft is summarized as follows: , in, For the cost of the voyage, For the sake of terrain, As a boundary cost, the range cost mainly considers the flight distance of the UAV from the starting point to the destination, and is proportional to the distance. If the total trajectory is... Composed of several waypoints, the total cost of the journey is: , Terrain cost primarily considers the threat posed by mountains during the drone's information acquisition mission. This cost constraint allows the drone to avoid obstacles during its operation. The terrain cost is as follows: , The boundary cost primarily considers ensuring that the UAV operates within a designated spatial area during the information acquisition task. The boundary cost is: , The amphibious drone is controlled to land near the search and rescue target according to the planned flight trajectory, enabling it to self-deploy for search and rescue. Through a high-precision deployment algorithm, it is deployed to the area around the person in the water. After deployment, the amphibious drone arrives at the rescue site and quickly approaches the person in the water. The person in the water can board the drone smoothly using the ergonomic concave surface and circular functional handle. After carrying the person in the water, the amphibious drone will quickly plan the optimal safe route to transfer the person to the nearest land or rescue platform, which greatly enhances the timeliness and reliability of maritime emergency rescue.

[0033] Please refer to the overall structural diagram of the self-deployed search and rescue drone. Figure 5 ,like Figure 5 As shown, the overall structure of the amphibious unmanned aerial vehicle (UAV) for self-deployed search and rescue includes a self-deployed search and rescue UAV 100 and an intelligent emergency control and communication center 200. For a structural diagram of the self-deployed search and rescue UAV, please refer to [link / reference needed]. Figure 6 ,like Figure 6 As shown, the self-deployed search and rescue drone includes a drone body 101, a hybrid power system 102, an energy management system 103, a camera 104, and a protective net 105. The drone body 101 adopts a concentric circle lifebuoy shape, with ergonomic concave surfaces 101a and ring-shaped functional handles 101b designed around its edge. The hybrid power system 102 includes an aerial propulsion unit 102a and a water propulsion unit 102b. The aerial propulsion unit 102a uses a dual-redundant brushless motor and a propeller with optimized aerodynamic design, enabling continuous and stable operation under complex water and weather conditions, providing strong power support for the drone's search and rescue operations. The water propulsion unit 102b uses an omnidirectional vector jet pump or a servo-controlled thruster, supporting high-speed aerial cruising and water surface operations. Flexible and mobile, to achieve Omnidirectional propulsion ensures omnidirectional movement in water, guaranteeing smooth transitions between aerial and surface power modules and reducing delays in rescue missions. Energy management module 103 utilizes a 4100mAh three-cell LiPo battery pack as the main power source, integrating a wireless charging module as auxiliary power. It incorporates multiple modules including current limiting protection, soft start, power monitoring, and multi-channel step-down circuits to ensure stable, safe, and long-lasting system operation. The wireless charging receiver is mounted around the bottom buffer pad, employing a small I-shaped ferrite core with a receiving coil wound on the outside. Ferrite is also placed below the transmitting coil, forming the main magnetic circuit. Since the receiving coil is wound on the ferrite at the receiving end, when alternating magnetic flux passes through the ferrite at the receiving end, it also passes through the receiving coil, thus achieving wireless power transmission. Camera 104 includes a visible light high-definition camera and an infrared camera radar. For a structural diagram of the intelligent emergency control and communication center, please refer to [link / reference needed]. Figure 7 ,like Figure 7 As shown, the intelligent emergency control and communication center 200 includes a communication and navigation module 201, a drone wireless charging module 202, a multi-dimensional monitoring module 203, and a detachable quick-release charging pile interface 204. The communication and navigation module 201 integrates GPS / RTK high-precision positioning, a data transmission radio, and a 4G / 5G communication link, enabling real-time data interaction and centimeter-level positioning with the drone platform. It supports 4G / 5G, Wi-Fi, and Bluetooth communication methods. Through these communication methods, information exchange between the drone and the charging pile can be achieved. For example, the drone can transmit its battery status and location information to the charging pile, and the charging pile can also provide feedback on charging parameters and operating status to the drone. It also facilitates remote monitoring and control of the charging pile and drone's operating status by a remote monitoring terminal, improving the intelligence and manageability of the charging process. The drone wireless charging module 202 is equipped with a Type-C interface and a Micro... The USB interface can adapt to the charging needs of different types of drones, ensuring the convenience of wireless charging while providing an alternative to wired charging for drones, and also ensuring the feasibility of charging in special circumstances such as wireless charging failure; the multi-dimensional monitoring module 203 includes temperature monitoring, humidity monitoring and wind speed monitoring units, which can monitor the temperature, humidity and wind speed of the drone charging environment in real time; the detachable quick-release charging pile interface 204 adopts a plug-in design, which is convenient for shore base stations or mother ships to install bases and directly and quickly replace charging piles.

[0034] In summary, the amphibious UAV self-drop search and rescue system provided by this invention includes a UAV self-drop search and rescue module and an intelligent emergency control module. The UAV self-drop search and rescue module is used to collect image data and point cloud data of the search and rescue area in real time according to the search and rescue mission, and transmit the image data and point cloud data to the intelligent emergency control module. The intelligent emergency control module is used to identify the search and rescue target based on the image data and point cloud data, track the identified search and rescue target, obtain the target pose, plan the flight trajectory of the UAV self-drop search and rescue module based on the target pose and the obtained pose of the UAV self-drop search and rescue module, and control the UAV self-drop search and rescue module to land near the search and rescue target based on the planned flight trajectory, thereby realizing amphibious UAV self-drop search and rescue and improving the timeliness and reliability of emergency rescue.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A self-deployed search and rescue system for an amphibious unmanned aerial vehicle (UAV), characterized in that, Includes a drone self-deployment search and rescue module and an intelligent emergency control module; The drone self-deployment search and rescue module is used to collect image data and point cloud data of the search and rescue area in real time according to the search and rescue mission, and transmit the image data and point cloud data to the intelligent emergency control module. The intelligent emergency control module is used to identify the search and rescue target based on the image data and point cloud data, track the identified search and rescue target, obtain the target pose, plan the flight trajectory of the UAV self-deployed search and rescue module based on the target pose and the obtained pose of the UAV self-deployed search and rescue module, and control the UAV self-deployed search and rescue module to land near the search and rescue target based on the planned flight trajectory, so as to realize the self-deployed search and rescue of the air and water amphibious UAV.

2. The air-sea amphibious unmanned aerial vehicle (UAV) self-deployment search and rescue system according to claim 1, characterized in that, The drone self-deployed search and rescue module includes a multimodal camera unit and a communication unit; The multimodal camera unit is used to collect image data and point cloud data of the search and rescue area in real time according to the search and rescue mission; The communication unit is used to transmit the image data and point cloud data to the intelligent emergency control module.

3. The air-sea amphibious unmanned aerial vehicle (UAV) self-deployment search and rescue system according to claim 2, characterized in that, The UAV self-deployment search and rescue module also includes a composite lightweight airframe structure, a hybrid propulsion unit, an energy and management unit, and a propeller protection net; The composite lightweight body structure is in the shape of a concentric circle lifebuoy. The edge of the concentric circle lifebuoy includes an ergonomic concave surface for supporting the search and rescue target and a ring-shaped functional handle for the search and rescue target to grip. The hybrid propulsion unit is used to switch between high-speed aerial cruising and agile maneuvering on the water. The energy and management unit is used to provide energy supply for the multimodal camera unit, communication unit, and hybrid propulsion unit; The propeller protection net is used to provide safety protection for the drone's propellers.

4. The self-deployed search and rescue system for amphibious unmanned aerial vehicles according to claim 2, characterized in that, The intelligent emergency control module includes an intelligent analysis unit, a trajectory planning unit, and a control unit; The intelligent analysis unit is used to perform search and rescue target detection on the image data using a deep learning model, obtain the category, detection box, and image texture features of the search and rescue target, extract the point cloud geometric features of the search and rescue target based on the point cloud data, match the point cloud geometric features with the image texture features to determine the search and rescue target and the search and rescue target pose, and use Kalman filtering to track the search and rescue target to obtain the target pose. The network architecture of the deep learning model is the YOLOv11n model. The trajectory planning unit is used to plan the flight trajectory of the UAV self-deployment search and rescue module based on the target pose and the acquired pose of the UAV self-deployment search and rescue module. The control unit is used to control the UAV self-deployment search and rescue module to land near the search and rescue target based on the planned flight trajectory, so as to realize the self-deployment search and rescue of the air and water amphibious UAV.

5. The self-deployed search and rescue system for amphibious unmanned aerial vehicles according to claim 4, characterized in that, The YOLOv11n model includes a backbone network, a neck network, a DyHead layer, and a detection head. The intelligent analysis unit is specifically used for: The image data is subjected to multi-scale feature extraction through the backbone network to obtain first-scale features, second-scale features and third-scale features. The first-scale feature, the second-scale feature, and the third-scale feature are fused using the neck network to obtain fused image features; The fused image features are enhanced through the attention mechanism of the DyHead layer, wherein the attention mechanism includes a scale-aware attention mechanism, a spatial-aware attention mechanism, and a task-aware attention mechanism. The enhanced image features are analyzed by a detection head to perform category prediction, bounding box regression, and target prediction, thereby obtaining the target's category, detection box, and image texture features.

6. The self-deployed search and rescue system for amphibious unmanned aerial vehicles according to claim 4, characterized in that, The loss function of the deep learning model is: ， ， ， ， ， ， ， in, The loss function of the deep learning model. The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. For distance loss, For angle loss, For the channel height, For channel width, For the first One detection box, For the first One detection box, For the predicted bounding box and the ground truth bounding box in Normalized distance in the direction, For the predicted bounding box and the ground truth bounding box in Normalized distance in the direction, The correlation coefficient for angle loss. The correlation coefficient for angle loss. This is for shape loss.

7. The self-deployed search and rescue system for amphibious unmanned aerial vehicles according to claim 4, characterized in that, The intelligent analysis unit is also used for: Calculate the cosine similarity between the point cloud geometric features and the image texture features; When the cosine similarity is greater than or equal to a preset similarity threshold, a correspondence between point cloud geometric features and image texture features is established. The point cloud geometric features corresponding to the image texture features are fused to determine the search and rescue target and its pose.

8. The self-deployed search and rescue system for amphibious unmanned aerial vehicles according to claim 4, characterized in that, The trajectory planning unit is specifically used for: The starting point and ending point of the UAV self-deployed search and rescue module are determined based on the target pose and the pose of the UAV self-deployed search and rescue module. Based on the starting point and the ending point, a cubic B-spline smoothing curve is used to plan the flight trajectory of the UAV self-deployment search and rescue module, and the planned flight trajectory is obtained.

9. The self-deployed search and rescue system for amphibious unmanned aerial vehicles according to claim 4, characterized in that, The intelligent emergency control module also includes a communication and navigation unit, a wireless charging unit, a multi-dimensional monitoring unit, and a charging pile interface unit; The communication and navigation unit is used to perform real-time data interaction and centimeter-level positioning with the UAV self-deployment search and rescue module. The wireless charging unit is used to provide wireless charging for the UAV self-deployment search and rescue module. The multi-dimensional monitoring unit is used to monitor the temperature, humidity, and wind speed of the wireless charging unit; The charging pile interface unit is used to provide an installation interface for the communication and navigation unit, the wireless charging unit, and the multi-dimensional monitoring unit.

10. A self-deployed search and rescue method for an amphibious unmanned aerial vehicle (UAV), characterized in that, This is achieved through the self-deployed search and rescue system for amphibious unmanned aerial vehicles as described in any one of claims 1-9, including: Based on the UAV self-deployment search and rescue module, image data and point cloud data of the search and rescue area are collected in real time according to the search and rescue mission, and the image data and point cloud data are transmitted to the intelligent emergency control module. The intelligent emergency control module identifies the search and rescue target based on the image data and point cloud data, tracks the identified target, obtains the target pose, plans the flight trajectory of the UAV self-deployed search and rescue module based on the target pose and the obtained pose of the UAV self-deployed search and rescue module, and controls the UAV self-deployed search and rescue module to land near the search and rescue target based on the planned flight trajectory, realizing the self-deployed search and rescue of the air and water amphibious UAV.

Images