An intelligent unmanned search and rescue system and method based on an improved PV-RCNN target detection algorithm

Abstract

The application belongs to the field of maritime search and rescue, and particularly relates to an intelligent unmanned search and rescue system and method based on an improved PV-RCNN target detection algorithm. In view of the problem that the sparse convolution mode in the traditional PV-RCNN algorithm cannot take into account the size of the receptive field and the convolution speed, the application proposes an improved PV-RCNN target detection algorithm. The method adopts a sparse convolution mode that adaptively changes the size of the receptive field, and also classifies the point cloud data of the detected target, so that the detection boat carrying the PV-RCNN algorithm can adaptively adjust its own angle on the water surface to obtain more abundant ship body semantic information, and improve the detection accuracy of the target. The intelligent search and rescue system of the application improves the target searching capability by fusing the advantages of the laser radar and the improved PV-RCNN target detection algorithm, can autonomously perceive, analyze and detect near the search and rescue target location, has a high detection and recognition speed, and improves the rescue speed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of maritime search and rescue, and particularly relates to an intelligent unmanned search and rescue system and method based on an improved PV-RCNN target detection algorithm. BACKGROUND

[0002] In recent years, unmanned ship technology and application in China have gradually been valued and continuously developed. For the rescue of unmanned ships that are out of order or stranded on the water surface, traditional unmanned ship rescue on the water surface consumes a large amount of manpower, and the route is planned by humans, so the rescue cost is high and the rescue work lacks systematicness. Therefore, an intelligent system is urgently needed to implement the emergency rescue task of unmanned ships in daily testing or related events. The rescue system needs to meet the navigation control function of the search and rescue ship on the predetermined route with short-distance target recognition task in all-weather and weak scenes. The existing three-dimensional target detection algorithm includes the PV-RCNN algorithm, but the PV-RCNN algorithm has a problem of sparse convolution mode. Common sparse convolution includes sub-manifold sparse convolution and regular sparse convolution. The former can only extract features in a specific size of receptive field, and the receptive field is limited, which cannot process disconnected features. Although the latter has a large enough fixed receptive field, the calculation amount is large and the GPU memory consumption is high. SUMMARY

[0003] In view of the problems and deficiencies in the prior art, the purpose of the present application is to provide an intelligent unmanned search and rescue system and method based on an improved PV-RCNN target detection algorithm.

[0004] To achieve the purpose of the application, the technical scheme adopted by the present application is as follows:

[0005] The present application provides a method for detecting water surface unmanned ships based on an improved PV-RCNN algorithm, comprising the following steps:

[0006] (1) Scanning the water surface scene to be detected by a laser radar to obtain original point clouds; performing voxelization processing on the obtained original point clouds to form a plurality of voxels, taking the average value of the features of the original point clouds in each voxel as the voxel feature of the voxel, and then extracting side view features of the point clouds by a Multi-View method to obtain a side view feature map;

[0007] (2) performing adaptive sparse convolution processing on the voxel features obtained in step (1) to obtain information exchange voxel features; performing downsampling operation on the information exchange voxel features to obtain multi-layer features of different scales;

[0008] (3) performing feature compression on the features obtained in step (2) in the multi-layer features with the highest sampling multiple in height to obtain an overhead feature map;

[0009] (4) performing down-sampling and up-sampling operations on the overhead feature map and the side-view feature map to obtain overhead feature maps and side-view feature maps of different scales, splicing the overhead feature maps and the side-view feature maps of the same scale to obtain spliced feature maps of multiple scales, and splicing the spliced feature maps of the multiple scales to obtain a final spliced feature map;

[0010] (5) generating a proposal box by using the anchor box method on the final spliced feature map;

[0011] (6) performing set abstraction operation on the highest sampling multiple in the multiple layers of features obtained in step (2), increasing the weight of the foreground points therein, and finally performing farthest point sampling on the foreground points to obtain key points;

[0012] (7) using the RoI grid pooling module to summarize the key points in step (6) into the features in the proposal box with multiple receptive fields to classify and regress the features, and complete the target detection task.

[0013] According to the method, preferably, the operation of step (2) of performing adaptive sparse convolution processing on the voxel features obtained in step (1) is that the number of point clouds in the voxel features is compared with a set threshold, and when the number of point clouds in the voxel features is greater than the set threshold, the voxel features are taken as information exchange voxel features. The set threshold is determined by the sparsity of the point cloud; the sparser the point cloud, the smaller the set threshold.

[0014] According to the method, preferably, the process of step (1) of scanning the water surface scene to be detected by using the laser radar to obtain the original point cloud is that the center axis of the bow to the stern of the water surface unmanned ship is taken as a standard line, the connecting line of the laser radar and the center of the unmanned ship body is taken as a detection line, and when the angle between the detection line and the standard line is 20°-160°, the point cloud information detected by the laser radar is defined as the side point cloud of the unmanned ship. The point cloud information in other angle ranges is the bow or stern point cloud; when the laser radar detects the stern or bow of the unmanned ship, the detection boat can adaptively move to a proper position to extract side data.

[0015] According to the method, preferably, the weight of the foreground points in step (6) is increased by using the following formula:

[0016]

[0017] wherein is the foreground point weight, is a three-layer multilayer perception machine, and f is the feature of the i-th layer at the position of the foreground point p.

[0018] According to the method, preferably, the downsampling operation in step (2) is a 1x, 2x, 4x, 8x downsampling operation on the voxel features obtained in step (1), to obtain multi-layer features of different scales.

[0019] According to the method, preferably, the manner of splicing the multi-scale spliced feature maps in step (4) is channel splicing.

[0020] According to the method, preferably, the process of scanning the water surface scene to be detected by the laser radar to obtain the original point cloud in step (1) is as follows:

[0021] The second aspect of the application provides an intelligent unmanned search and rescue system based on a laser radar and a PV-RCNN detection method, which comprises a main console and a search and rescue boat, and a communication module, a navigation control module, a sensing module, a rescue module and an industrial control module are arranged on the search and rescue boat; the main console is used for receiving a search and rescue task and acquiring the size and GPS information of a searched boat, and the size and GPS information of the searched boat are communicated to the search and rescue boat to update the position information of the searched boat in real time; the navigation control module is used for guiding the search and rescue boat to autonomously fly to an accident site; the sensing module is used for confirming the identity of the searched boat and detecting the position of the searched boat; the sensing module comprises a laser radar, the laser radar acquires point cloud data by emitting a laser beam, and the point cloud data is processed by the method of the first aspect to complete the task of detecting the position of the searched boat.

[0022] The rescue module comprises a control case and an electromagnet, the searched boat is pulled out by the electromagnet and then sails back to the search and rescue boat in parallel, and the industrial control console is used for coordinating the task allocation and information transmission of the searched boats and the search and rescue boat.

[0023] According to the intelligent unmanned search and rescue system, preferably, the navigation control module comprises a GPS, an IMU and an electronic map, and a composite positioning method of GPS positioning, inertial positioning and electronic map matching positioning is adopted.

[0024] The third aspect of the application provides a search and rescue method of the intelligent unmanned search and rescue system of the second aspect, which comprises the following steps:

[0025] S1: the main console sends a command to the search and rescue boat through the communication module;

[0026] S2: the search and rescue boat receives the task information and leaves the shore;

[0027] S3: the search and rescue boat goes to the target position through the navigation control module, and the specific position of the searched boat is detected by the laser radar and the detection method based on the PV-RCNN algorithm of the first aspect;

[0028] S4: the search and rescue boat is connected with the searched boat through the rescue module;

[0029] S5: the search and rescue boat and the searched boat plan a route for returning through the navigation control module, and obstacle avoidance is performed through the laser radar during the returning;

[0030] S6: the search and rescue boat and the searched boat are landed.

[0031] According to the search and rescue method, preferably, the specific process of step 1 is that when an unmanned boat on the water surface has an accident, the searched boat sends a distress information, and the information mainly includes GPS information and basic shape model information of the searched boat; after the main control table receives the distress information of the searched boat, the searched boat is sent a command through a communication module, and the GPS information of the searched boat is given and updated in real time at a certain frequency.

[0032] According to the search and rescue method, preferably, the specific process of step 2 is that the search and rescue boat receives the real-time updated searched boat information, and the search and rescue boat equipped with a laser radar senses the surrounding environment and then leaves the shore.

[0033] According to the search and rescue method, preferably, the specific process of step 3 is:

[0034] 3.1 After the search and rescue boat leaves the shore, the navigation control module makes the search and rescue boat autonomously fly to the vicinity of the target GPS position according to the GPS position of the searched boat and in combination with the obstacle avoidance system of the laser radar;

[0035] 3.2 If the GPS signal of the searched boat is good, the search and rescue boat directly approaches the searched boat through the acquired GPS information of the searched boat and senses and detects the specific position of the searched boat through the laser radar and the detection method based on the PV-RCNN algorithm in the first aspect; if the searched boat loses the GPS signal near the shore, the search and rescue boat approaches the searched boat according to the last GPS information of the searched boat, senses through the laser radar after approaching, and performs target detection through the detection method based on the PV-RCNN algorithm in the first aspect to determine the specific position of the searched boat.

[0036] According to the search and rescue method, preferably, in step 3.1, the fusion positioning of the GPS and the IMU is realized by means of the extendable Kalman filtering method provided by the ROS platform after the search and rescue boat leaves the shore.

[0037] Compared with the prior art, the present application has the following beneficial effects:

[0038] (1) In view of the problem that the sparse convolution mode in the traditional PV-RCNN algorithm cannot balance the receptive field size and the convolution speed, the application proposes an improved PV-RCNN target detection algorithm. The method uses a sparse convolution mode that adaptively changes the receptive field size, which can balance the information exchange between adjacent point clouds and the speed requirement. At the same time, the application also classifies the point cloud data of the detected target, so that the detection boat equipped with the PV-RCNN detection algorithm can adaptively adjust its own angle on the water surface to obtain more rich ship body semantic information, and improve the detection accuracy of the target.

[0039] (2) The application also provides a system and method for searching and rescuing unmanned surface vehicles that have failed or run aground on the water. The search and rescue system has high flexibility and strong real-time performance when performing water search and rescue tasks. The application improves the target searching capability by combining the advantages of laser radar and the improved PV-RCNN target detection algorithm, and can autonomously perceive, analyze and detect near the target search location, with high detection and recognition speed, and improves the rescue speed. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 is a schematic diagram of the intelligent unmanned surface vehicle search and rescue system of the application;

[0041] Figure 2 is a GPS / IMU fusion framework diagram of the application;

[0042] Figure 3 is a laser radar perception target diagram of the search and rescue boat when approaching the searched boat;

[0043] Figure 4 is a point cloud diagram of the search and rescue boat and the searched boat returning side by side;

[0044] Figure 5 is a projection diagram of the laser radar scanning the shore point cloud when the search and rescue boat approaches the shore. DETAILED DESCRIPTION

[0045] In order for those skilled in the art to more clearly understand the technical solutions of the application, the technical solutions of the application will be described in detail below with specific examples.

[0046] Example 1

[0047] A method for detecting unmanned surface vehicles based on an improved PV-RCNN algorithm, comprising the following steps:

[0048] (1) Scan the water surface scene to be detected with lidar to obtain the original point cloud; perform voxelization on the obtained original point cloud to form multiple voxels, take the average value of the original point cloud features in each voxel as the voxel feature of that voxel, and then extract the side view features of the point cloud through the Multi-View method to obtain the side view feature map.

[0049] The process of obtaining the original point cloud is as follows: Using the centerline from the stern to the bow of the unmanned surface vessel (USV) as the standard line, and the line connecting the lidar and the center of the hull as the detection line, the angle between the detection line and the standard line is defined as 90°–180° when the lidar detects point clouds at the bow or bow + one side, and 0°–90° when the lidar detects point clouds at the stern or stern + one side. Therefore, the angle corresponding to the lidar detecting point clouds on the side is 20°–160°.

[0050] The results of collecting water surface point cloud data in this embodiment are shown in Table 1:

[0051] Table 1. Point cloud data of water surface

[0052]

[0053] As shown in Table 1, the semantic information of the LiDAR for the detected vessel is poor when only the bow or stern point clouds are available, while the side point clouds provide richer semantic information. Based on the attitude of the detected vessel exposed to the detection vessel, and given the detection vessel's relatively free navigation on the water surface, the detection vessel can adaptively adjust its bow angle to detect more side information of the detected vessel. This ensures richer semantic information even when the detected vessel is at a greater distance, resulting in higher recognition accuracy.

[0054] For example, in this embodiment, when the detection boat identifies the detected point cloud as the bow and stern as shown in Table 1, the detection boat's trajectory and relative position with the detected boat are changed to identify the detected boat. Assuming the detected boat is directly in front of the detection boat, i.e., when the stern of the detected boat is detected, the detection boat will move to the right front to an appropriate position to ensure that the side can be detected. At this time, there is richer semantic information, making the detection more robust.

[0055] (2) Adaptive sparse convolution processing is performed on the voxel features obtained in step (1). The specific operation is as follows: the number of point clouds in the voxel features is compared with a set threshold. When the number of point clouds in the voxel features is greater than the set threshold, the voxel features are used as information exchange voxel features. For example, if the original convolution kernel is 3*3*3, and the kernel of this size contains 27 neighborhood boxes, it is determined whether there are point clouds in these voxels, and a threshold is determined according to the distribution of the number of point clouds, thereby determining whether to count the voxel as an information exchange voxel. The threshold is determined according to the sparsity of the point cloud; the sparser the point cloud, the smaller the threshold. Finally, this invention can achieve selective dilation within the receptive field to perform adaptive information exchange.

[0056] Then, the voxel features of information exchange are downsampled to obtain multi-layer features at different scales;

[0057] (3) Compress the features obtained from the multi-layer features in step (2) using the highest sampling factor in terms of height to obtain a top-view feature map;

[0058] Since the detection only involves the bow and stern, the side and top views of the unmanned surface vessel (USV) have larger areas than the front and rear views, resulting in more complete semantic information. This invention adds the side and top view information as features to the network for training and uses the BEV (height-compressed) and Multi-View method to extract the semantic information from the top and side views, further improving the detection accuracy of the search and rescue vessel.

[0059] (4) Perform downsampling and upsampling operations on the top view feature map obtained in step (3) and the side view feature map obtained in step (1) to obtain top view feature maps and side view feature maps of different scales. Then, stitch the top view feature maps and side view feature maps of the same scale together to obtain stitched feature maps of multiple scales. Finally, stitch the stitched feature maps of multiple scales together by channel to obtain the final stitched feature map.

[0060] (5) Generate suggestion boxes from the final stitched feature map using the anchor box method;

[0061] (6) Perform set abstraction operation on the multi-layer features obtained in step (2) using the highest sampling multiple, and then apply the formula Increase the weight of the foreground points, among which Let A() be the weight of the foreground points, and A() be a three-layer multilayer perceptron, where f is the feature of the i-th layer at position p of the foreground point. Finally, the farthest point of the foreground point is sampled to obtain the key points.

[0062] (7) Use the RoI grid pooling module to aggregate the key points in step (6) into a RoI grid with multiple receptive fields to classify and regress the features in the proposal box, thus completing the object detection task.

[0063] Example 2

[0064] This embodiment provides an intelligent unmanned search and rescue system. When an unmanned surface vessel malfunctions or runs aground on the water surface of a test site, the main control station receives a rescue mission and dispatches a search and rescue vessel to the target sea area to search for and rescue the malfunctioning unmanned surface vessel.

[0065] like Figure 1 As shown, this invention provides an intelligent unmanned surface vessel (USV) search and rescue system, comprising a main control console and the USV itself. The USV is equipped with a communication module, a flight control module, a sensing module, a rescue module, and an industrial control module. The main control console receives search and rescue missions and acquires the size and GPS information of the USV being searched, and communicates the USV's GPS information to the USV to update its location in real time. The flight control module navigates the USV autonomously to the target GPS location (or accident site). The sensing module confirms the identity of the USV and detects its specific location. The rescue module pulls the USV out and returns it safely alongside the USV. The industrial control console coordinates the task allocation and information transmission between multiple USVs and the USV.

[0066] The flight control module includes GPS, IMU, and electronic map, and adopts a composite positioning method that combines GPS positioning, inertial positioning, and electronic map matching positioning.

[0067] The perception module includes a lidar, which acquires point cloud data by emitting laser beams to the surrounding area. It mainly scans and acquires the size, shape, and distance of the rescue vessel to construct a real-time 3D environment. Then, the improved PV-RCNN algorithm in Example 1 is used to process the point cloud data to confirm the location of the rescue vessel.

[0068] Example 3

[0069] A search and rescue method for an intelligent unmanned surface vessel search and rescue system includes the following steps:

[0070] Step 1: The main control console sends commands to the search and rescue boat via the communication module;

[0071] The specific process is as follows: When an unmanned vessel is involved in an accident on the water, the vessel being searched sends out a distress signal. This signal mainly includes the GPS information and basic shape and model information of the vessel being searched. After receiving the distress signal from the vessel being searched, the main control console sends a command to the vessel through the communication module, providing the GPS information of the vessel being searched and updating it in real time at a certain frequency.

[0072] Step 2: The search and rescue boat receives the mission information and departs from shore;

[0073] The specific process is as follows: the search and rescue boat receives real-time updates on the information of the boat being searched and rescued, and the search and rescue boat equipped with lidar senses the surrounding environment and then leaves the shore.

[0074] Step 3: The search and rescue boat travels to the vicinity of the target location via the navigation control module and detects the specific location of the search and rescue boat using the improved PV-RCNN algorithm in Example 1;

[0075] The specific process is as follows:

[0076] 3.1 After the search and rescue boat leaves the shore, the navigation control module uses the GPS position of the searched boat and the obstacle avoidance system of the lidar to enable the search and rescue boat to autonomously navigate to the vicinity of the target GPS position.

[0077] During this process, some extreme situations may be encountered, such as narrow river channels and sections with overpasses on the riverbed. The GPS signal of the search and rescue vessel may be lost, unstable, or have a low refresh rate (typically 10 Hz), causing the unmanned vessel to collide with the shore and cause an accident. In this embodiment, the GPS latitude and longitude are used as input signals to the IMU, allowing the IMU to navigate through narrow river channels by measuring certain parameters. Furthermore, LiDAR SLAM can be used to construct an image of the surrounding environment, which, combined with the IMU's measurement, enables the vessel to pass through narrow and winding river channels.

[0078] The specific process of integrating GPS and IMU is as follows: Figure 2 As shown:

[0079] An IMU (Inertial Measurement Unit) typically consists of three accelerometers and three gyroscopes, capable of measuring acceleration and angular velocity in three directions, respectively, using a... x ,a y ,a z ,w x ,w y ,w z It indicates that GPS (Global Positioning System) is used to acquire GPS positioning information by carrying a GPS receiver on the unmanned surface vessel (USV). The information that can be obtained includes accuracy, latitude, and altitude.

[0080] Because GPS operates at frequencies below 10Hz, which are insufficient for real-time performance, while IMUs operate at frequencies above 50Hz, GPS and IMU may receive data at different times. Therefore, when GPS has not yet generated data and only IMU data is received, the IMU data is processed by the motion model to obtain a predicted state X(ˇ)k. Simultaneously, this predicted state is transmitted back to the motion model for the next prediction step. Alternatively, when GPS data is generated, the previously generated predicted state is fused with the received GPS location information to obtain a corrected state X(^)k. Finally, this corrected state is transmitted back to the motion model for subsequent motion model predictions.

[0081] The state vector of the motion model consists of three parts: position, velocity, and direction. These parameters are updated using the following method:

[0082] Position p k The formula is In the formula, fk-1 is the measurement value of the IMU, and C ns It is used for coordinate transformation of IMU measurements;

[0083] Assuming the object is undergoing uniformly accelerated linear motion, the velocity Vk is given by the formula Vk = Vkk. k =V k-1 +Δt(C ns f k-1 -g)

[0084] Direction q k The formula is q k =Ω(q(w) k-1 Δt))q k-1 In the formula q k-1 Represented by quaternions;

[0085] Basic principles such as Figure 2 As shown, GPS / IMU fusion positioning is achieved by using the callable extended Kalman filter method provided by the ROS platform.

[0086] 3.2 If the GPS signal of the rescue vessel is good, the rescue vessel will approach the vessel directly by acquiring the GPS information of the rescue vessel and perceive the specific location of the rescue vessel by lidar.

[0087] If the rescue vessel loses its GPS signal near the shore, its movement range is limited because it may have run aground. Therefore, the rescue vessel can approach the shore based on the last GPS information of the vessel being searched, and then use lidar to detect it. Figure 3 The lidar imagery of the search and rescue vessel as it approaches the vessel being searched. Figure 3The improved PV-RCNN algorithm was used to detect targets and determine the exact location of the search and rescue vessel.

[0088] After obtaining the suggestion box of the rescue vessel through detection, the relative distance d and direction α of the rescue vessel can be calculated. The GPS position information G2(long2,lat2) of the rescue vessel can be calculated using the GPS information G1(long1,lat1) of the rescue vessel.

[0089] The calculation method is as follows:

[0090]

[0091] long2=long1+d*sinα / [ARC*cos(lat1)*2π / 360]

[0092] lat2=lat1+d*cosα / (ARC*2π / 360)

[0093] Step 4: The search and rescue boat connects to the boat being searched via the rescue module;

[0094] The specific process is as follows: The search and rescue boat is equipped with a control box and an electromagnet. This device is installed on the side of the search and rescue boat, and the search and rescue boat attracts the boat being searched through this device.

[0095] Step 5: The search and rescue vessel and the vessel being searched return via the navigation control module, using lidar for obstacle avoidance during the process. The return process in Step 5 is similar to that in Step 3. The point cloud image of the search and rescue vessel and the vessel being searched returning side by side is shown below. Figure 4 As shown;

[0096] Step 6: The search and rescue boat and the boat being searched and rescued dock.

[0097] The specific process is as follows: Figure 5 The image shown is a projection of the point cloud on the shore scanned by the lidar when the search and rescue boat docks. The boundary line of the shore base is extracted based on the point cloud data and obtained through projection transformation. The route is planned by judging the shore base, and the route maintains a certain distance from the shore base and docks parallel to it. When docking, the return time of different laser emission points is different due to the influence of water flow speed and boat speed, which will cause distortion of the point cloud. Therefore, it is necessary to combine the speed and acceleration information of the IMU and the position information of the GPS to correct the real-time position of the boat to achieve a more stable docking.

[0098] The above embodiments are specific implementations of the present invention, but the implementation of the present invention is not limited to the above embodiments. Any other combination, change, modification, substitution, or simplification that does not exceed the design concept of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for detecting unmanned surface vessels based on an improved PV-RCNN algorithm, characterized in that, Includes the following steps: (1) Scan the water surface scene to be detected with lidar to obtain the original point cloud; perform voxelization on the obtained original point cloud to form multiple voxels, take the average value of the original point cloud features in each voxel as the voxel feature, and then extract the side view features of the point cloud through the Multi-View method to obtain the side view feature map. (2) Adaptive sparse convolution processing is performed on the voxel features obtained in step (1) to obtain information exchange voxel features; the information exchange voxel features are downsampled to obtain multi-layer features of different scales. (3) Compress the features obtained from the multi-layer features in step (2) using the highest sampling factor in terms of height to obtain the top view feature map; (4) Perform downsampling and upsampling operations on the top view feature map and the side view feature map to obtain top view feature maps and side view feature maps of different scales. Then, stitch the top view feature maps and side view feature maps of the same scale together to obtain a stitched feature map of multiple scales. Finally, stitch the stitched feature maps of multiple scales together to obtain the final stitched feature map. (5) Generate suggestion boxes from the final stitched feature map using the anchor box method; (6) Perform set abstraction operation on the multi-layer features obtained in step (2) using the highest sampling multiple, then increase the weight of the foreground points, and finally sample the farthest point of the foreground points to obtain the key points; (7) Use the RoI grid pooling module to aggregate the key points in step (6) into a RoI grid with multiple receptive fields to classify and regress the features within the proposal box, thus completing the target detection task.

2. The method according to claim 1, characterized in that, The operation of adaptive sparse convolution processing of the voxel features obtained in step (1) in step (2) is as follows: the number of point clouds in the voxel features is compared with a set threshold. When the number of point clouds in the voxel features is greater than the set threshold, the voxel features are used as information exchange voxel features.

3. The method according to claim 2, characterized in that, The process of scanning the water surface scene to be detected with lidar to obtain the original point cloud in step (1) is as follows: taking the centerline from the bow to the stern of the unmanned surface vessel as the standard line, and the line connecting the lidar and the center of the unmanned surface vessel as the detection line, the angle between the detection line and the standard line is defined as 20°. o ~160 o At that time, the point cloud information detected by the lidar is the side point cloud of the unmanned surface vessel.

4. The method according to claim 3, characterized in that, In step (6), the weight of the foreground points is increased using the following formula: in As for the weight of the foreground attractions, ( ) is a three-layer multilayer perceptron. f It is the front attraction. p Position below i Characteristics of the layer.

5. The method according to claim 4, characterized in that, The downsampling operation is to perform 1x, 2x, 4x and 8x downsampling operations on the voxel features obtained in step (1) to obtain multi-layer features of different scales.

6. The method according to claim 5, characterized in that, The method of stitching the feature maps of multiple scales in step (4) is to stitch them together by channel.

7. An intelligent unmanned search and rescue system based on lidar and PV-RCNN detection method, characterized in that, The system includes a main control console and the search and rescue boat itself, which is equipped with a communication module, a navigation control module, a sensing module, a rescue module, and an industrial control module. The main control console is used to receive search and rescue missions and obtain the size and GPS information of the rescue vessel, and communicate the size and GPS information of the rescue vessel to the rescue vessel to update the location information of the rescue vessel in real time. The flight control module is used to guide the search and rescue boat to autonomously navigate to the accident site. The perception module is used to confirm the identity of the rescue vessel and detect its location; the perception module includes a lidar, which acquires point cloud data by emitting a laser beam and processes the point cloud data using the method described in claim 1 to complete the task of detecting the location of the rescue vessel. The rescue module search and rescue boat is pulled out by an electromagnet and then returns safely to shore in parallel with the search and rescue boat. The industrial control console is used to coordinate the task allocation and information transmission between multiple search and rescue vessels.

8. The search and rescue method of the intelligent unmanned search and rescue system according to claim 7, characterized in that, Includes the following steps: S1: The main control console sends commands to the search and rescue boat via the communication module; S2: The search and rescue boat receives mission information and departs from shore; S3: The search and rescue boat travels to the target location via the navigation control module and detects the specific location of the search and rescue boat using the method described in claim 1; S4: The search and rescue boat connects to the boat being searched and rescued via a rescue module; S5: The search and rescue boat and the boat being searched return to port along the route planned by the navigation control module, during which they use lidar to sense the environment and avoid obstacles. S6: The search and rescue boat and the boat being searched and rescued dock.

9. The search and rescue method according to claim 8, characterized in that, The specific process of step S3 is as follows: 3.1 After the search and rescue boat leaves the shore, the navigation control module uses the GPS position of the searched boat and the obstacle avoidance system of the lidar to enable the search and rescue boat to autonomously navigate to the vicinity of the target GPS position. 3.2 If the GPS signal of the vessel being searched is good, the search and rescue vessel will approach directly using the acquired GPS information of the vessel being searched and will perceive and detect the specific location of the vessel being searched using the detection method described in claim 1; if the vessel being searched loses its GPS signal near the shore, the search and rescue vessel will approach based on the last GPS information of the vessel being searched, and after approaching, will perceive the target using lidar and use the detection method described in claim 1 to detect the target and determine the specific location of the vessel being searched.

10. The search and rescue method according to claim 9, characterized in that, In step 3.1, after the search and rescue boat leaves the shore, the extended Kalman filter method provided by the ROS platform is used to achieve GPS and IMU fusion positioning.

Images