A search and rescue watercraft for all terrain

Abstract

A kind of all-terrain search and rescue water surface boat, including water surface boat body, driving mechanism, multiple distributed cameras, core computing power control board, shore-based monitoring platform, data transmission module and remote control handle. Multiple cameras are arranged circumferentially around the body, and multiple video streams are synchronously collected under hardware synchronous trigger. The core computing power control board is built-in lightweight deep learning target detection model, which detects rescue targets in real time and parallel inference on multiple video streams, and calculates target direction and distance through multi-view fusion positioning. According to the detection result, driving control instruction is generated to drive the execution mechanism, and autonomous approach is realized. The shore-based monitoring platform receives and displays multiple videos and target superimposed information through HDMI to WIFI, and switches between autonomous approach mode and manual remote control mode without disturbance. The invention realizes all-terrain coverage, fast target identification and autonomous approach, and effectively improves the efficiency and safety of water surface rescue by combining real-time monitoring and dual-mode remote control.

Description

Technical Field

[0001] This invention relates to the field of unmanned surface vessels (USVs), specifically to a surface vessel for all-area search and rescue. Background Technology

[0002] With the rapid development of water tourism, water operations, and the marine economy, surface activities are becoming increasingly frequent. Simultaneously, maritime accidents such as ship collisions, people falling overboard, and natural disasters are also on the rise, leading to a continuous increase in the demand for surface rescue. Traditional surface rescue methods primarily rely on manually operated rescue vessels or helicopters for searches, which suffer from slow response times, limited search areas, and high personnel safety risks in harsh environments, making it difficult to meet the practical needs of rapid search and rescue across large areas of water.

[0003] In recent years, unmanned surface vehicles (USVs) have gradually become important equipment for surface search and rescue due to their advantages such as maneuverability, low cost, and ability to replace personnel in dangerous areas. USVs can autonomously navigate along preset routes and can be remotely controlled via a remote control center, showing broad application prospects in target search, environmental monitoring, and emergency rescue. Extensive research both domestically and internationally has focused on the environmental perception, autonomous navigation, and remote control of USVs, with some solutions attempting to achieve environmental perception and target detection by equipping them with cameras, radar, and navigation sensors. However, existing surface rescue vessels still have significant shortcomings in target search and identification, specifically in the following aspects.

[0004] (i) The field of view of a single or a small number of cameras is limited, making it difficult to achieve full coverage.

[0005] Most existing surface rescue boats are equipped with only a single camera, resulting in a limited field of view and making it difficult to achieve omnidirectional, blind-spot-free near-field target search. While some solutions attempt to expand the field of view by installing panoramic or 360° cameras, a single panoramic camera has inherent limitations in resolution, target detail capture, and long-distance recognition capabilities, failing to meet the dual requirements of wide field of view coverage and high-precision target identification. Other solutions employ a distributed multi-camera layout, such as an automated anchoring vision assistance system for unmanned surface vessels based on 3D panoramic surround view. This system uses cameras installed in four directions on the unmanned surface vessel to collect surrounding video information and stitch it together to form a panoramic view. However, these solutions are primarily geared towards collision avoidance and anchoring assistance applications. Their multi-video processing focuses on image stitching and panoramic display, lacking the ability to simultaneously retrieve and identify targets from multiple video streams, making them unsuitable for the rapid target discovery and location requirements of active search and rescue scenarios.

[0006] (ii) There is a lack of effective synchronous acquisition and fusion processing capabilities for multiple video streams.

[0007] While some multi-camera solutions can acquire multiple video streams, they lack effective mechanisms for simultaneous acquisition and fusion processing. Images acquired by different cameras are difficult to keep strictly synchronized in time, resulting in unclear spatiotemporal correspondences of the same target in different camera feeds. This makes it impossible to effectively correlate and fuse target information from multiple video streams, thus hindering the improvement of target detection probability and positioning accuracy. Furthermore, most existing solutions separate video acquisition from target recognition processing—cameras are only responsible for image acquisition, while image recognition is performed by shore-based equipment. This means that target recognition results cannot be fed back to the shipboard control system in real time, making it difficult to form a closed-loop control chain of "perception-decision-execution."

[0008] (iii) The image transmission link and the control link are independent of each other and lack efficient coordination.

[0009] In existing unmanned surface vessel (USV) systems, image transmission and remote control command transmission often rely on independent communication links. Image data is typically transmitted back to the shore via a dedicated video transmission channel, while control commands are transmitted to the USV via a separate wireless link. This separate link design not only increases the system's complexity and cost but also makes it difficult to efficiently coordinate image transmission and control command transmission. Although operators can view the video feed transmitted from the USV through the shore-based monitoring platform, the issuance of control commands and the transmission of images occur independently on different channels, resulting in time delays and asynchrony, which affects the real-time performance and user experience of remote control.

[0010] (iv) There is a lack of flexible mode switching mechanism between autonomous identification and manual remote control.

[0011] Existing surface rescue boats either rely entirely on manual remote control or use preset programs for autonomous navigation, lacking a mechanism for flexible switching between autonomous identification and approach and manual intervention. In autonomous mode, the system often only cruises along a preset path and cannot dynamically adjust its course based on real-time target perception. In remote control mode, operators need to maintain high levels of attention to monitor the screen and operate the handle for extended periods, which can easily lead to fatigue and delayed response. Switching between the two modes usually requires complex system restarts or mode resets, failing to achieve seamless and dynamic switching and making it difficult to balance the efficiency of automated search with the flexibility of manual intervention.

[0012] In addition, several solutions related to water rescue have been disclosed in the prior art, but each has different technical limitations. For example, Chinese invention patent application CN117799799A discloses a "rescue method for an unmanned surface vessel used for garbage disposal," which uses a camera to identify and locate accidents by setting up a sound acquisition and recognition system and an image acquisition and recognition system on the unmanned vessel. However, this solution is mainly aimed at auxiliary rescue functions in garbage disposal scenarios, using a single or a small number of cameras to identify accident points, failing to achieve full field-of-view coverage, and lacking the ability to simultaneously retrieve multiple video streams and fuse core computing power, making it unsuitable for full-area target search in large-scale water rescue scenarios. Summary of the Invention

[0013] To address the shortcomings of existing technologies, this invention provides a surface vessel for all-area search and rescue, which overcomes the deficiencies of existing technologies, achieves full-area coverage, rapid target identification and autonomous approach, and, combined with shore-based real-time monitoring and dual-mode remote control, effectively improves the efficiency and safety of surface rescue.

[0014] To achieve the above objectives, the present invention provides the following technical solution:

[0015] A surface vessel for all-area search and rescue includes:

[0016] A surface vessel body,

[0017] A drive mechanism is provided at the stern of the surface vessel body. The controlled end of the drive mechanism is connected to the control output end of the core computing power control board, and is used to receive drive control commands from the core computing power control board and perform propulsion power output and navigation direction turning operations according to the drive control commands.

[0018] Multiple cameras are arranged in a circumferentially around the surface vessel body. The field of view of each camera together forms a field of view covering the area around the surface vessel body. Each camera is used to synchronously acquire images of the water surface area around the surface vessel body and transmit the acquired multi-channel video streams to the core computing power control board.

[0019] A core computing power control board is mounted inside the surface vessel. The input terminals of the core computing power control board are connected to the output terminals of each of the cameras. The core computing power control board is used to receive multiple video streams from multiple cameras. The core computing power control board is equipped with an automatic detection and recognition driving algorithm. The automatic detection and recognition driving algorithm processes the multiple video streams in parallel to detect and identify rescue targets in an omnidirectional area. Based on the orientation information of the identified rescue targets relative to the surface vessel, it generates driving control commands and outputs the driving control commands to the driving mechanism.

[0020] A data transmission module, mounted on the surface vessel, is electrically connected to the core computing power control board and is used to establish a wireless data communication link between the core computing power control board and the shore-based monitoring platform and remote control handle.

[0021] A shore-based monitoring platform establishes a wireless communication link with a data transmission module via HDMI to WIFI. The shore-based monitoring platform is used to receive and display image data from multiple cameras forwarded by the data transmission module in real time.

[0022] A remote control handle is provided, which is wirelessly connected to the core computing power control board via a data transmission module. The remote control handle is used to send remote control commands to the core computing power control board and receive execution status information fed back by the core computing power control board.

[0023] The core computing power control board has an autonomous approach mode and a manual remote control mode. In the autonomous approach mode, the core computing power control board autonomously generates drive control commands based on the recognition results of the automatic detection and recognition drive algorithm. In the manual remote control mode, the core computing power control board controls the drive mechanism in response to the remote control commands.

[0024] Preferably, the multiple cameras adopt a distributed array layout, and the field of view of each camera is configured such that the fields of view of adjacent cameras at least partially overlap, and the fields of view of all cameras are stitched together to form a panoramic field of view covering a 360° range around the body of the boat on the water surface; the horizontal field of view of each camera is not less than 120°.

[0025] Preferably, the core computing power control board is equipped with a hardware clock synchronization unit, which is used to add a hardware timestamp to each channel image frame when the video streams collected by each camera are input to the core computing power control board; the core computing power control board performs frame-level synchronization and alignment of multiple video streams according to the hardware timestamps of each channel.

[0026] Preferably, the automatic detection and recognition driving algorithm has a built-in target detection model based on a deep learning neural network. The target detection model runs on the neural network processing unit built into the core computing power control board and performs parallel inference operations on multi-channel video streams. The target detection model is implemented with a lightweight network structure and outputs the bounding box coordinates, category confidence score, and category label of the target for each input frame image. The core computing power control board is set with a confidence threshold. Only when the category confidence score is not lower than the confidence threshold is the target determined to be a valid rescue target.

[0027] Preferably, the automatic detection and recognition driving algorithm further includes an azimuth calculation function. When the target detection model detects the same rescue target in multiple video frames at the same time, the core computing power control board extracts the center coordinates of the bounding box of the target in each camera's image. Combining the pre-calibrated parameters of each camera, it solves the azimuth angle φ and distance d of the rescue target relative to the surface vessel body through triangulation or perspective projection methods. The core computing power control board calculates the deflection angle δ and the rotational speed n of the drive mechanism based on the deviation Δφ between the current bow direction and the azimuth angle φ, and generates a corresponding PWM control signal to output to the drive mechanism to drive the surface vessel body to sail toward the target.

[0028] Preferably, the shore-based monitoring platform includes a display terminal and a WIFI receiving module; the core computing power control board combines multiple video streams collected by multiple cameras into a single multi-view composite video, which is output to the data transmission module via the HDMI interface of the core computing power control board. The HDMI to WIFI encoding unit in the data transmission module encodes the multi-view composite video and then wirelessly transmits it to the shore-based monitoring platform via the WIFI protocol; the shore-based monitoring platform displays images from each camera's perspective in real time on the display terminal in a multi-view split format, allowing operators to grasp the overall situation of the rescue site; simultaneously, after detecting a rescue target, the core computing power control board overlays the target's location information onto the screen of the corresponding channel of the shore-based monitoring platform, and marks the target's distance and azimuth parameters on the screen.

[0029] Preferably, the remote control handle is equipped with a mode switching switch. When the mode switching switch is activated, the remote control handle sends a mode switching command to the core computing power control board through the data transmission module. In autonomous approach mode, the joystick operation of the remote control handle is disabled or used only as an auxiliary input. The core computing power control board autonomously plans the trajectory and controls the drive mechanism to execute based on the recognition results of the automatic detection and recognition drive algorithm. In manual remote control mode, the joystick displacement of the remote control handle is directly mapped to the propulsion speed command and steering angle command of the drive mechanism. The operator makes remote control decisions based on the real-time images displayed on the shore-based monitoring platform and issues drive control commands through the remote control handle.

[0030] Preferably, the core computing power control board is configured to, when operating in autonomous approach mode, if no valid rescue target is detected within a preset number of consecutive frames, control the drive mechanism to drive the surface vessel along a preset search trajectory; if a valid rescue target is detected during the cruise, automatically switch to target tracking mode and control the surface vessel to sail towards the target; when the distance between the surface vessel and the target decreases to within a preset rescue distance threshold, the core computing power control board controls the drive mechanism to stop propulsion and waits for manual confirmation or triggering the release of rescue equipment.

[0031] Preferably, the data transmission module is used to establish a two-way wireless data communication link between the core computing power control board, the shore-based monitoring platform, and the remote control handle. The data transmission module operates at a frequency of 2.4 GHz or 5.8 GHz, and the communication distance is not less than 500 meters. The data transmission module simultaneously carries the downlink transmission link for image data and the uplink transmission link for control commands, so that image transmission and remote control command transmission are carried in the same wireless channel through time division multiplexing and / or frequency division multiplexing.

[0032] Preferably, the surface vessel body has multiple sealed compartments, each including an equipment compartment, a buoyancy compartment, and a battery compartment; the equipment compartment is a watertight sealed compartment, and the core computing power control board and data transmission module are housed within the equipment compartment; each camera is respectively installed on the bow and the top of the side bulwarks of the surface vessel body, and each camera has an independent waterproof housing.

[0033] This invention provides a surface vessel for all-area search and rescue, offering the following advantages: It employs three or more cameras arranged in a circumferentially distributed configuration around the vessel, forming a panoramic field of view covering 360° of the vessel's perimeter. Furthermore, the fields of view of adjacent cameras overlap, fundamentally overcoming the inherent limitations and blind spots of single-channel cameras. Based on this, a synchronization trigger signal is generated by the hardware clock synchronization unit of the core computing power control board, controlling simultaneous exposure of each camera. A hardware timestamp is added to each frame during image acquisition, achieving strict frame-level synchronization of multiple video streams. This makes the spatiotemporal correspondence of the same target in different camera images clearly traceable, providing a precise data foundation for multi-view fusion positioning of the target. When the same target appears simultaneously in the fields of view of multiple cameras, triangulation or perspective projection methods are used to calculate the target's azimuth and distance relative to the surface vessel based on the target's pixel coordinates in each channel image and the camera's calibration parameters. This achieves a precise mapping from "image detection" to "spatial positioning," significantly improving target positioning accuracy and reliability compared to existing single-camera solutions or simple image stitching schemes.

[0034] By directly deploying the automatic detection and recognition algorithm on the ship's core computing control board, and utilizing the built-in neural network processing unit (KPU) to perform parallel inference operations on multiple video streams, target detection and recognition are completed in real time on the ship, eliminating the need to upload images to the cloud or shore-based servers for processing. This fundamentally eliminates the impact of network transmission latency on real-time performance. This allows the surface vessel to automatically turn towards the target immediately upon detection, without waiting for manual remote control commands, significantly shortening the response time from target discovery to approaching rescue.

[0035] By using a single data transmission module to simultaneously carry both downlink video image transmission and uplink control command transmission, bidirectional transmission of video streams and control commands is achieved within the same wireless channel using time-division multiplexing or frequency-division multiplexing. Compared to existing technologies where image transmission and control links are independent, this invention simplifies the system architecture and reduces hardware costs and the probability of communication conflicts. More importantly, transmitting video data and control commands on the same channel naturally ensures the time correlation between image return and the issuance of drive control commands. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of this invention or the prior art will be briefly introduced below.

[0037] Figure 1 The overall system architecture and connection diagram of this invention;

[0038] Figure 2 Flowchart of the target detection and autonomous control program of this invention;

[0039] Explanation of the labels in the diagram:

[0040] 1. Surface vessel body; 2. Drive mechanism; 3. Multiple cameras; 4. Core computing power control board; 5. Shore-based monitoring platform; 6. Data transmission module; 7. Remote control handle. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0042] Example 1, as Figures 1 to 2 As shown, a surface vessel for all-area search and rescue includes: a surface vessel body 1, a drive mechanism 2, multiple cameras 3, a core computing power control board 4, a shore-based monitoring platform 5, a data transmission module 6, and a remote control handle 7.

[0043] The surface vessel hull 1 adopts a catamaran or monohull deep-V hull structure, possessing excellent surface navigation stability and resistance to wind and waves. The surface vessel hull 1 features a multi-compartment sealed structure, with each compartment isolated from the others by watertight bulkheads, enabling autonomous floating and capsizing resistance. Even if one compartment accidentally takes on water, the remaining compartments can still provide sufficient reserve buoyancy to maintain the surface vessel's normal floating attitude. The surface vessel hull 1 achieves motion control through a drive mechanism 2, which carries various components and provides the platform for water navigation.

[0044] The drive mechanism 2 is installed at the stern of the surface vessel body 1 and electrically connected to the core computing power control board 4. It is used to receive drive control commands to realize the power output and steering control of the surface vessel. In this embodiment, the drive mechanism 2 includes: a thruster, installed at the center of the stern of the surface vessel body, for providing navigation power; a steering servo, installed in conjunction with the thruster, for changing the navigation direction; and a drive control board, connected to the PWM output terminal of the core computing power control board, for converting the PWM control signal output by the core computing power control board into the speed control voltage of the thruster and the angle control voltage of the steering servo. The thruster is preferably a brushless DC motor driven propeller thruster, which has the characteristics of fast response and high efficiency; the steering servo is preferably a waterproof digital servo, with an angle control accuracy of not less than 1°.

[0045] Multiple cameras 3 are distributed and installed around the hull of the surface vessel 1 for simultaneous multi-view image acquisition of near targets. All cameras 3 are connected to the core computing power control board 4 via CSI cables, transmitting the acquired multi-channel video streams to the core computing power control board 4. In this embodiment, the cameras are wide-angle cameras with a horizontal field of view of not less than 120°, a resolution of not less than 1920×1080 pixels, a frame rate of not less than 30fps, and automatic exposure and automatic white balance functions to adapt to changes in water surface lighting conditions. Each camera is equipped with an independent waterproof housing with a protection rating of not less than IP67.

[0046] The core computing power control board 4 is fixed inside the surface vessel body 1. It is used to receive multiple video streams from multiple cameras 3, and uses an automatic detection and recognition driving algorithm to synchronously search and identify nearby targets. Based on the recognition results, it outputs driving control commands to the driving mechanism 2. At the same time, it interacts with external devices through the data transmission module 6.

[0047] The shore-based monitoring platform 5 connects to the data transmission module 6 via an HDMI-to-WIFI converter, and then establishes communication with the core computing power control board 4. This allows for real-time reception and display of image data collected by multiple cameras 3, enabling real-time image transmission from multiple cameras and monitoring of the rescue site. The shore-based monitoring platform 5 includes a display terminal and a WIFI receiving module. The display terminal is preferably a portable laptop or tablet computer, equipped with dedicated monitoring software capable of presenting images from each camera's perspective in a multi-screen split format.

[0048] The data transmission module 6 is mounted on the surface vessel hull 1 and provides a wireless data communication link between the core computing power control board 4, the shore-based monitoring platform 5, and the remote control handle 7, with a communication distance of no less than 500m. The data transmission module 6 preferably adopts a dual-band WIFI communication module, with operating frequencies including 2.4GHz and 5.8GHz bands, and supports IEEE 802.11ac or higher protocols.

[0049] The remote control handle 7 is connected to the core computing power control board 4 via the data transmission module 6. It sends remote drive control commands to the core computing power control board 4 to remotely control the surface vessel 1 and receives execution status information from the core computing power control board 4. The remote control handle 7 is a handheld remote controller equipped with a joystick, buttons, and status indicator lights. It integrates a WIFI communication module, which communicates directly with the data transmission module 6.

[0050] Example 2 is basically the same as Example 1, except for the layout of the multiple cameras 3.

[0051] Multiple cameras 3 are arranged in a distributed array and installed at the front and left and right sides of the surface vessel 1, forming a 360° panoramic field of view. The specific layout is as follows:

[0052] One wide-angle camera is mounted at the front end, fixed to the bow deck or bow mast of the surface vessel via a bracket. The lens optical axis points directly forward, with a pitch angle tilted downwards by 5° to 15°, for forward search and target locking. The front-end camera preferably uses a large-aperture wide-angle lens to obtain good imaging results in low-light conditions, with a horizontal field of view of not less than 120° and a vertical field of view of not less than 70°.

[0053] One wide-angle camera is mounted on each of the port and starboard sides of the surface vessel, respectively, fixed to the top of the bulwark on the port and starboard sides or on a dedicated camera bracket. The optical axes of the lenses are pointed to the port and starboard sides, respectively, with the pitch angle tilted downwards by 5° to 15° to expand the lateral search range and detect near-field targets on both sides of the vessel. The field-of-view parameters of the cameras on both sides are consistent with those of the front camera to ensure consistent image resolution and imaging quality in all directions.

[0054] The optical centers of all cameras are located in the same horizontal plane, the height of which is determined based on the hull's draft and the expected search distance, typically ranging from 0.5m to 1.5m. There is at least a 10° overlap between the fields of view of adjacent cameras to ensure no blind spots at the field of view boundaries and to provide sufficient parallax information for multi-view target fusion and localization.

[0055] With the above arrangement, the three cameras together form a horizontal field of view of 180°–270°, covering the front, port, and starboard sides. In a variant, to further achieve stern coverage, a fourth wide-angle camera can be added to the stern of the surface vessel to monitor the waters behind, forming a complete 360° panoramic field of view. This stern camera is fixed to the stern of the surface vessel by a bracket, with its lens optical axis pointing directly aft, and its parameters are consistent with those of the front camera.

[0056] Each camera is calibrated before leaving the factory. The calibration includes the camera's intrinsic parameter matrix, distortion coefficients, and extrinsic parameters (rotation matrix and translation vector) of each camera relative to the coordinate system of the surface vessel. The calibration method adopted is the Zhang Zhengyou calibration method. After calibration, the distortion correction parameters and projection matrix of each camera are stored in the memory of the core computing power control board 4 for subsequent image correction and orientation calculation.

[0057] Example 3: Based on Example 1 or Example 2, this example further details the hardware implementation scheme of the core computing power control board 4 and the synchronous acquisition mechanism of multiple video streams.

[0058] The core computing power control board 4 uses an embedded AI development board with an integrated AI accelerator as the main processing control board. For example, it can employ a SoC (System-on-Chip) solution equipped with a KPU (Knowledge Processing Unit) AI accelerator. This SoC integrates a dual-core or quad-core CPU processor and a dedicated KPU neural network accelerator. The core computing power control board 4 is equipped with at least 2GB of LPDDR4 RAM and at least 16GB of eMMC storage space for storing the operating system, algorithm model weights, and video cache data. The specific hardware interface configuration of the core computing power control board 4 is as follows:

[0059] Camera interface: Connects multiple cameras via MIPI CSI interface 3, supports multi-channel MIPI CSI-2 protocol, and the data transmission rate per channel is not less than 1.5Gbps;

[0060] Video output interface: Provides MIPI DSI or HDMI interface for outputting composite video images to data transmission module 6 for wireless transmission;

[0061] Control output interface: Provides multiple PWM (Pulse Width Modulation) output interfaces for outputting steering and propulsion control signals to the drive mechanism;

[0062] Communication interfaces: Provides general-purpose interfaces such as UART, SPI, and I2C for communication with data transmission modules and other peripherals;

[0063] To achieve frame-level synchronization of multiple video streams, the core computing power control board 4 is equipped with a hardware clock synchronization unit. Its synchronization acquisition and frame synchronization processing flow is as follows:

[0064] (1) Synchronous trigger signal generation: The hardware clock synchronization unit of the core computing power control board 4 generates a synchronous trigger signal (square wave pulse), which is output to the external trigger input terminal of each camera through the GPIO pin. The frequency of the synchronous trigger signal is consistent with the camera frame rate (e.g., 30Hz), and the rising edge triggers each camera to start the exposure of a frame of image simultaneously.

[0065] (2) Synchronous exposure and image acquisition: After receiving the rising edge of the synchronization trigger signal, each camera starts exposure and reads out image data simultaneously. Since all cameras share the same trigger signal, the image frames acquired by each camera at the same time are strictly aligned in time, and the deviation of the exposure start time is less than 1 microsecond.

[0066] (3) Hardware timestamp annotation: When the image data collected by each camera is transmitted to the core computing power control board 4 through the MIPI CSI interface, the CIM (Camera Interface Module) module in the core computing power control board 4 automatically adds a hardware timestamp at the Frame Start time of each frame image to record the precise time when the frame image arrives at the core computing power control board. This hardware timestamp is based on the system clock of the core computing power control board and has an accuracy of microseconds.

[0067] (4) Frame-level synchronization alignment: When reading image frames from each channel, the driver of the core computing power control board 4 performs frame-level synchronization alignment based on the hardware timestamp of each frame. Specifically, for image frames from different cameras within the same time window (e.g., ±5 milliseconds), they are determined to be a synchronized frame group at the same acquisition time for use by subsequent target detection and fusion localization algorithms. If a channel fails to arrive within the time window due to transmission delay or other reasons, the alignment process is waited for the frame to arrive before proceeding, or the frame is discarded to ensure the time consistency of the synchronized frame group.

[0068] By combining hardware synchronization triggering with hardware timestamp annotation, the core computing power control board 4 can ensure strict synchronization of multiple video streams at the frame level, providing a reliable spatiotemporal correspondence for subsequent multi-view target fusion and positioning.

[0069] Example 4 is basically the same as Example 1, except that the automatic detection and recognition driving algorithm is implemented in a specific way.

[0070] The automatic detection and recognition algorithm is mounted on the main processor of the core computing power control board 4. It uses a lightweight target detection model based on deep learning neural networks to perform parallel inference on the multi-channel image data collected by the multiple cameras 3. The specific implementation steps of the algorithm are as follows:

[0071] (I) Construction and Deployment of Target Detection Model

[0072] The automatic detection and recognition algorithm incorporates a built-in object detection model, employing a lightweight convolutional neural network architecture (such as YOLOv5s, YOLOv8s, or MobileNet-SSD). YOLOv5n or YOLOv8n is preferred as the baseline network. Lightweight processing, including model pruning and INT8 quantization, enables real-time operation on the KPU AI accelerator of the core computing power control board 4. The object detection model takes a 640×640 pixel image as input and outputs the bounding box coordinates (x, y, y) of the target. center ,y center The model includes width, height, category confidence score, and category label. Before deployment, the model requires the following training and transformation process:

[0073] (1) Dataset construction: Collect images of water surface targets under different lighting conditions, weather conditions, and water environments, including common target categories in rescue scenarios such as people falling into the water (human body), life rings, small boats, and buoys. Mark the bounding boxes and categories of the targets in the images to construct a dedicated water surface rescue target detection dataset.

[0074] (2) Model training: The YOLO network is trained on the PC using the dataset. Data augmentation strategies (including random flipping, rotation, color jittering, mosaic enhancement, etc.) are used to improve the generalization ability of the model. The training continues until the average accuracy (mAP) of the model on the validation set reaches the preset requirements.

[0075] (3) Model conversion and quantization: The trained PyTorch model is exported to ONNX format, and then the ONNX model is converted to KModel format using the K230 model conversion tool. At the same time, INT8 quantization is performed to reduce the model size and improve the inference speed. The quantized model weight file is stored in the eMMC flash memory of the core computing power control board.

[0076] (4) Model loading: When the system is powered on and initialized, the core computing power control board 4 loads the model weight file in KModel format from the eMMC flash memory to the KPU memory to complete the deployment of the target detection model.

[0077] (ii) Multi-channel parallel inference mechanism

[0078] Because the KPU AI accelerator on the core computing power control board 4 supports tensor parallel computing, the automatic detection and recognition driving algorithm processes the video streams from the three cameras using a pipelined parallel inference approach. Specifically:

[0079] a. The video streams from the three cameras are cached in different memory areas of the core computing power control board;

[0080] b. The algorithm submits three frames of images at the same time to the KPU for inference sequentially or in parallel according to the synchronized frame group after frame synchronization alignment;

[0081] c.KPU performs object detection inference on each frame of image and outputs the object detection results in that frame of image (object bounding box coordinates, class confidence score and class label).

[0082] d. The detection results of each channel are stored separately for use by the subsequent target fusion and localization module.

[0083] At a frame rate of 30fps, the total inference time of the core computing power control board for three 640×640 resolution images does not exceed 30 milliseconds, which meets the real-time processing requirements.

[0084] (III) Confidence threshold determination

[0085] The core computing power control board 4 is set with a confidence threshold (preferably 0.5). The target detection model assigns a category confidence score (range 0-1) to each target output in each frame of the image. The core computing power control board 4 compares the confidence score of each target with the confidence threshold: if the confidence score of a target is not lower than the confidence threshold, the target is determined to be a valid rescue target; if the confidence score of a target is lower than the confidence threshold, the target is determined to be a false alarm or a low-confidence target and is not responded to. Furthermore, the algorithm uses non-maximum suppression (NMS) to merge overlapping detection boxes. The NMS threshold is preferably set to 0.45 to eliminate duplicate detections of the same target.

[0086] (iv) Multi-view target fusion positioning

[0087] When the target detection model detects the same rescue target in multiple video frames at the same time (i.e., the same target appears in the field of view of two or three cameras at the same time), the orientation calculation function of the automatic detection and recognition driving algorithm performs the following fusion localization steps:

[0088] a. Target Association: Based on the target bounding box positions, target categories, and confidence scores detected in each channel, and combined with the field-of-view overlap relationship of each camera, target boxes corresponding to the same physical target in different channels are associated and matched. Specifically, when the target categories detected in two channels are the same and the positions of the bounding boxes in their respective images satisfy the epipolar constraint condition, they are determined to be the same target.

[0089] b. Extracting target center coordinates: For the same target that has been successfully associated, the core computing power control board 4 extracts the center coordinates (u_i) of the target's bounding box in the images of each camera. i ,v i ), where i represents the camera channel number (i=1,2,3).

[0090] c. Determining the target azimuth and distance: Combining the pre-calibrated parameters of each camera (including intrinsic matrix, distortion coefficients, and extrinsic matrix, i.e., the installation position and attitude angle of each camera relative to the surface vessel), the azimuth φ and distance d of the rescue target relative to the surface vessel are determined using a multi-view geometric method. In this embodiment, a triangulation method is employed: Based on the pixel coordinates of the same target in the images of two cameras, the spatial three-dimensional coordinates (X, Y, Z) of the target are calculated using the triangulation principle of binocular vision, and then converted into the azimuth φ (with the bow direction as 0° and clockwise as positive) and straight-line distance d of the target relative to the vessel. When the target appears simultaneously in the fields of view of three cameras, the least squares method is used to fuse the three sets of observation data to improve positioning accuracy.

[0091] If the target is detected by only a single camera, the azimuth and distance of the target are estimated using a monocular perspective projection method, based on the target's position in the image, the camera calibration parameters, and the height constraint of the horizontal plane where the surface vessel is located (assuming the target is at the water's surface height). Although the accuracy of monocular estimation is lower than that of multi-view fusion positioning, it can still provide an effective azimuth reference when the target is within the field of view of a single camera.

[0092] (v) Heading calculation and drive control command generation

[0093] The core computing power control board 4, based on the output target azimuth angle φ and distance d, and combined with the current heading information of the submarine (provided by the inertial measurement unit (IMU) or electronic compass), executes the following heading calculation and drive control steps:

[0094] a. Calculate heading deviation: Calculate the deviation angle Δφ=φ-θ between the current bow angle θ and the target azimuth angle φ, and normalize Δφ to the range of [-180°, 180°].

[0095] b. Calculate the deflection angle: Calculate the deflection angle δ of the drive mechanism based on the deviation angle Δφ. In this embodiment, a proportional control strategy is adopted: δ = Kp × Δφ, where Kp is the proportional coefficient (preferably 0.8 to 1.2), and the value range of the deflection angle δ is limited to [-45°, 45°] to prevent excessive steering from causing instability of the hull.

[0096] c. Calculate propulsion speed: Calculate the propulsion speed of the drive mechanism based on the target distance d. When the target distance is far (e.g., d > 50m), use a high-speed cruising speed (e.g., 3-5 knots); when the target distance is moderate (e.g., 10m < d ≤ 50m), use a medium-speed approach speed (e.g., 2-3 knots); when the target distance is close (e.g., d ≤ 10m), decelerate to a low speed (e.g., below 1 knot) and prepare to stop. Speed ​​control also adopts a proportional control strategy: n = Kv × d, where n is the propeller speed command, Kv is the speed proportional coefficient, and a maximum speed limit is set to prevent overspeeding.

[0097] d. Generate PWM control signals: The core computing power control board 4 converts the deflection angle δ and the thruster speed n into the duty cycle of the PWM signal, respectively, and generates the steering control PWM signal and the propulsion control PWM signal.

[0098] e. The PWM output interface of the core computing power control board 4 is output to the drive control board of the drive mechanism 2. The drive control board receives the PWM commands from the core computing power control board 4: the steering PWM signal drives the servo motor or steering motor to rotate the steering device to the target deflection angle δ; the propulsion PWM signal drives the propeller motor controller to adjust the propeller speed to the target speed n. The drive mechanism 2 responds to the steering and propulsion commands simultaneously, driving the surface vessel body 1 to accelerate and approach the target along the optimal course. Real-time steering and propulsion control of the surface vessel body 1 is achieved, completing the autonomous navigation control towards the target.

[0099] The execution result of the drive mechanism 2 is fed back to the core computing power control board 4 in real time through the encoder or Hall sensor, forming a closed-loop control to ensure that the actual deflection angle and propulsion speed are consistent with the command value.

[0100] Example 5 is basically the same as Example 1, except that it is different from the complete control process of the autonomous approach mode of the core computing power control board 4.

[0101] The control flow of the core computing power control board 4 in autonomous approach mode is as follows:

[0102] Step S1: System power-on initialization

[0103] After the core computing power control board 4 is powered on, it performs the following initialization operations: loads the weight file of the target detection model into the KPU memory; initializes the drivers of each camera and sets the image acquisition parameters (resolution, frame rate, exposure time, white balance, etc.); initializes the PWM output interface and sets the initial values ​​of PWM frequency and duty cycle; initializes the WIFI communication module and establishes a data link with the shore-based monitoring platform 5 and the remote control handle 7; reads the current heading data of the ship's onboard inertial navigation module or electronic compass; and reads the preset search trajectory parameters (such as search area boundaries, cruising speed, etc.).

[0104] Step S2: Simultaneous data acquisition from multiple cameras

[0105] The hardware clock synchronization unit of the core computing power control board 4 generates a synchronization trigger signal to control the three cameras to simultaneously expose and acquire images. The images acquired by each camera are transmitted to the core computing power control board 4 through the MIPI CSI interface. When receiving images, the core computing power control board adds a hardware timestamp to each frame.

[0106] Step S3: Image Preprocessing

[0107] The core computing power control board 4 performs the following preprocessing operations on the images of each channel:

[0108] The three video streams are synchronized and aligned at the frame level based on the hardware timestamp to ensure that the three frames at the same time form a synchronized frame group; the images of each channel are scaled from the original resolution (e.g., 1920×1080) to the model input size (640×640); the images are normalized, mapping the pixel values ​​from [0, 255] to the range of [0, 1] or [-1, 1]; the normalized image data is converted into RGB tensor format, ready to be input into the object detection model.

[0109] Step S4: Object Detection Inference

[0110] The three preprocessed images are input sequentially or in parallel into the object detection model for inference. The object detection model performs forward propagation computation on the KPU AI accelerator, outputting the detection result for each image frame, including: the bounding box coordinates of the target (x, y, z). center ,y center (width, height), category confidence score, and category label.

[0111] Step S5: Determine whether a valid rescue target has been detected.

[0112] The core computing power control board 4 iterates through the detection results of each channel and compares the category confidence score of each target with a preset confidence threshold (e.g., 0.5):

[0113] If no valid rescue target with a confidence level exceeding the threshold is detected in any of the camera channels, proceed to step S6 (cruise search state).

[0114] If at least one camera channel detects a valid rescue target, proceed to step S7 (target tracking state).

[0115] Step S6: Cruise Search

[0116] When no valid rescue target is detected through any of the channels, the core computing power control board 4 controls the surface vessel 1 to cruise along a preset search path. The preset search path can be a bow-shaped scanning path, a spiral search path, or any path planned based on GPS coordinates. During the cruise, the core computing power control board 4 continuously executes steps S2 to S5, cyclically detecting targets. If a valid rescue target is detected during the cruise, it automatically switches to the target tracking state in step S7.

[0117] Step S7: Multi-view target fusion localization

[0118] When a valid rescue target is detected, the core computing power control board 4 performs multi-view target fusion positioning: it associates and matches the same target detected in different channels; extracts the center coordinates of the target's bounding box in each channel; and, in combination with the calibration parameters of each camera, solves the azimuth angle φ and distance d of the target relative to the surface vessel body through triangulation.

[0119] Step S8: Calculate the optimal heading and propulsion speed

[0120] Based on the deviation Δφ = φ - θ between the target azimuth angle φ and the current bow direction θ, the core computing power control board 4 calculates the optimal deflection angle δ = Kp × Δφ, and calculates the thruster speed n = Kv × d based on the target distance d. The deflection angle δ and the speed n are then converted into the duty cycle of the PWM control signal.

[0121] Step S9: Drive mechanism execution

[0122] The core computing power control board 4 sends steering control signals and propulsion control signals to the drive control board of the drive mechanism 2 via the PWM output interface. The drive control board converts the PWM signals into the steering angle control voltage of the steering servo and the speed control voltage of the propeller, driving the surface vessel body 1 to sail towards the target direction.

[0123] Step S10: Real-time video stream transmission

[0124] Throughout the autonomous approach process, the core computing power control board 4 continuously synthesizes the video streams from the three cameras into a single multi-view composite video (such as a three-screen display), outputs it via HDMI interface to the HDMI-to-WIFI encoding unit in the data transmission module 6, and then transmits it in real time to the shore-based monitoring platform 5 via WIFI protocol, allowing shore-based operators to remotely monitor the rescue site situation. Simultaneously, the core computing power control board 4 overlays the target detection results (detection frame, target category, confidence level, azimuth angle, distance, etc.) onto the corresponding channel's image, enabling operators to intuitively understand the target identification and positioning status.

[0125] Step S11: Execute repeatedly or the task ends.

[0126] If the target distance d decreases to within the preset rescue distance threshold (e.g., 2 meters), the core computing power control board 4 controls the drive mechanism 2 to stop propulsion, so that the surface vessel body stops near the target, waiting for the operator to manually confirm through the shore-based monitoring platform or trigger the automatic release of the rescue equipment.

[0127] If the target is lost during the tracking process (i.e., no valid target is detected for several consecutive frames), return to step S6, switch to cruise search state, and continue searching;

[0128] If a manual takeover command is received from the remote control handle 7 during the tracking process, the system will switch to manual remote control mode and hand over control to the operator.

[0129] Example 6 is basically the same as Example 1, except that the implementation method and function of the shore-based monitoring platform 5 are different.

[0130] The shore-based surveillance platform 5 includes a display terminal and a WIFI receiver module. The display terminal is preferably a portable laptop, tablet, or dedicated monitor, equipped with dedicated surveillance software. The WIFI receiver module is matched with the data transmission module 6 and supports dual-band WIFI communication of 2.4GHz and 5.8GHz.

[0131] (a) Real-time display of multiple screens

[0132] The core computing power control board 4 combines the three video streams collected by the three cameras 3 into a single multi-view composite video. The multi-view composite video is presented in a three-screen format: the left screen is the image from the port side camera, the middle screen is the image from the bow camera, and the right screen is the image from the starboard side camera. The three screens are displayed synchronously in time, allowing operators to simultaneously grasp the water situation in three directions: in front of the hull, to the port front, and to the starboard front, forming an intuitive perception of the overall situation at the rescue site.

[0133] The multi-view composite video is output to the data transmission module 6 via the HDMI interface of the core computing power control board 4. The HDMI-to-WIFI encoding unit in the data transmission module 6 encodes and compresses the multi-view composite video using H.264 or H.265 encoding, and then wirelessly transmits it to the shore-based monitoring platform 5 via the WIFI protocol. The WIFI receiving module of the shore-based monitoring platform 5 receives the video stream, decodes it using the monitoring software, and displays it in real time on the display terminal.

[0134] (ii) Overlay display of target detection results

[0135] After detecting a rescue target, the core computing power control board 4 overlays the target detection results onto the corresponding channel's screen in real time. Specifically:

[0136] On the channel screen where a target is detected, the target's position in the image is marked with a colored rectangle (detection box); above or next to the detection box, the target category label (such as "person", "lifebuoy", "boat") and confidence score (such as "0.92") are displayed; in a corner of the screen or near the target, the target's azimuth angle (such as "φ=35°") and distance (such as "d=18.5m") relative to the surface vessel are displayed.

[0137] When the same target appears in multiple channels at the same time, the detection box of the target in each channel is displayed in the same color, and a "multi-view lock" status mark is marked on the screen to facilitate operators to confirm the consistency of the target.

[0138] (III) Display of navigation status information

[0139] The monitoring software of the shore-based surveillance platform 5 also overlays the current navigation status parameters of the surface vessel onto the display interface, including but not limited to:

[0140] Current heading (e.g., "heading: 125°");

[0141] Current speed (e.g., "Speed: 2.3 knots");

[0142] Current operating mode ("Autonomous Approach Mode" or "Manual Remote Control Mode");

[0143] Target lock status ("Locked" or "Searching");

[0144] Device status information such as battery level and communication signal strength.

[0145] (iv) Operator interaction functions

[0146] The monitoring software of the shore-based surveillance platform 5 also provides the following interactive functions:

[0147] Mode switching confirmation: When the operator issues a mode switching command via the remote control handle 7, the monitoring software pops up a confirmation dialog box to prevent accidental operation.

[0148] Target Confirmation: In autonomous approach mode, when the surface vessel approaches the target to a preset distance, the monitoring software prompts the operator to confirm the target. The operator can trigger the automatic release of the rescue equipment by clicking the "Confirm" button on the screen or operating the confirmation button on the remote control handle.

[0149] Emergency Stop: The monitoring software is equipped with an emergency stop button. When an abnormal situation is detected, the operator can send an emergency stop command with one click to stop the surface vessel immediately.

[0150] Example 7 is basically the same as Example 1, except that the remote control handle 7 and the dual-mode switching mechanism are different.

[0151] (a) Hardware configuration of the remote control handle

[0152] The remote control handle 7 is a handheld wireless remote control, and its hardware configuration includes:

[0153] Joysticks: At least a left joystick and a right joystick. The left joystick is used to control the advance speed (push and pull) and the turning angle (swipe left and right), and the right joystick is used to control the camera's pitch angle.

[0154] Mode switching switch: A three-position toggle switch or button for switching between "autonomous approach mode" and "manual remote control mode";

[0155] Status indicator lights: used to indicate the current working mode (e.g., solid green indicates autonomous mode, solid red indicates manual mode), communication status (e.g., flashing indicates communication interruption) and battery level;

[0156] Confirm button: Used to confirm the target and trigger the release of rescue equipment in autonomous approach mode;

[0157] Emergency Stop Button: A separate large red button used to send an emergency stop command;

[0158] Built-in WIFI communication module: communicates directly with data transmission module 6, with an operating frequency of 2.4GHz or 5.8GHz.

[0159] (II) Dual-mode switching mechanism

[0160] The remote control handle 7 sends a mode switching command to the core computing power control board 4 via the data transmission module 6, achieving seamless switching between autonomous approach mode and manual remote control mode. The specific operating methods of the two modes are as follows:

[0161] Autonomous approach mode:

[0162] In autonomous approach mode, the core computing power control board 4 autonomously plans its trajectory and controls the drive mechanism 2 to execute actions based entirely on the recognition results of the automatic detection and recognition drive algorithm. Specifically:

[0163] The joystick operation signal of the remote control handle 7 is shielded or ignored by the core computing power control board 4 (except for the emergency stop button); the core computing power control board 4 autonomously executes the closed-loop control of "image acquisition → target detection → fusion positioning → heading calculation → drive control" according to the control process of embodiment five; the core computing power control board 4 continuously transmits the images of each camera, target detection results and recognition results back to the shore-based monitoring platform 5 for remote monitoring by the operator; the operator can understand the autonomous approach status of the surface vessel in real time through the shore-based monitoring platform 5, but does not directly intervene in the control (except in emergency situations).

[0164] Manual remote control mode:

[0165] In manual remote control mode, the remote control handle 7 gains complete control of the surface vessel. Specifically:

[0166] The core computing power control board 4 directly maps the joystick displacement of the remote control handle 7 to the propulsion speed command and steering angle command of the drive mechanism 2; the forward and backward push and pull of the left joystick controls the propulsion speed (push forward to accelerate, pull back to decelerate / reverse), and the left and right flick of the left joystick controls the steering angle (left flick for left turn, right flick for right turn); the core computing power control board 4 continues to perform target detection and recognition, and transmits the detection results and video stream to the shore-based monitoring platform 5 in a bypass manner for the operator's reference; the operator makes remote control decisions based on the real-time images displayed on the shore-based monitoring platform 5 and issues control commands through the remote control handle 7.

[0167] Disruptive design for mode switching:

[0168] Operators can switch between "autonomous approach mode" and "manual remote control mode" at any time by toggling the mode switch on the remote control handle 7. The switching process does not require interrupting system operation or restarting or re-initializing any hardware or software modules.

[0169] During the switching process, the core computing power control board 4 adopts a "bumper-free switching" strategy: when switching from manual remote control mode to autonomous approach mode, the core computing power control board 4 immediately starts the target search and target detection process with the current position and heading of the surface vessel as the initial state, without generating control jumps; when switching from autonomous approach mode to manual remote control mode, the core computing power control board 4 immediately stops the output of autonomous control commands, and instead maps the current joystick position of the remote control handle to drive control commands. When the joystick is in the middle position, the drive mechanism 2 maintains its current state, avoiding sudden acceleration or turning caused by the mismatch between the joystick position and the current control quantity.

[0170] (III) Priority Logic of Mode Switching

[0171] The core computing power control board 4 is equipped with priority logic for mode switching:

[0172] In any mode, the emergency stop button on the remote control handle 7 has the highest priority. Once pressed, the core computing power control board 4 immediately controls the drive mechanism to stop propulsion, regardless of the current mode.

[0173] In autonomous approach mode, if the target is detected to be lost or the target exceeds the preset safe distance threshold, the core computing power control board 4 automatically reduces the propulsion speed and controls the surface vessel to perform a search action in the predetermined area. At the same time, it sends a prompt signal to the shore-based monitoring platform and the remote control handle through the data transmission module, suggesting that the operator switch to manual remote control mode for manual search.

[0174] In manual remote control mode, if the communication link is interrupted for more than a preset time (e.g., 3 seconds), the core computing power control board 4 will automatically switch to autonomous approach mode and perform safety protection actions (e.g., decelerate and stop the boat) to prevent the surface boat from going out of control.

[0175] Example 8 is basically the same as Example 1, except that the communication implementation method of data transmission module 6 is different.

[0176] The data transmission module 6 is mounted on the surface vessel body 1 and is electrically connected to the core computing power control board 4. It is used to provide a two-way wireless data communication link between the core computing power control board 4, the shore-based monitoring platform 5, and the remote control handle 7.

[0177] (a) Hardware configuration

[0178] Data transmission module 6 adopts a dual-band WIFI communication module, and its hardware configuration includes:

[0179] WIFI chip: Supports IEEE 802.11ac or higher protocol, and operates in two frequency bands: 2.4GHz and 5.8GHz;

[0180] Power amplifier: Built-in power amplifier and low-noise amplifier, communication distance not less than 500 meters (open water).

[0181] Antenna: External high-gain omnidirectional antenna with a gain of not less than 5dBi;

[0182] HDMI to WIFI encoding unit: Built-in H.264 / H.265 hardware encoder, used to encode and compress the video signal input from HDMI and then transmit it via WIFI;

[0183] Interfaces: HDMI input interface (to receive video output from the core computing power control board), UART or SPI interface (to interact with the core computing power control board for control commands).

[0184] (ii) Data transmission link

[0185] Data transmission module 6 simultaneously supports the following two data transmission links:

[0186] a. Video data downlink transmission link (main link):

[0187] Transmission direction: from the surface vessel body (core computing power control board) to the shore base (shore-based monitoring platform);

[0188] Transmitted content: Multi-screen composite video (H.264 / H.265 encoded and compressed video stream);

[0189] Transmission protocol: RTSP or HTTP-FLV streaming media protocol;

[0190] Bandwidth requirement: Approximately 2 to 5 Mbps (depending on resolution and encoding parameters).

[0191] b. Control command uplink transmission link (secondary link):

[0192] Transmission direction: From shore-based (remote control handle / shore-based monitoring platform) → surface vessel body (core computing power control board);

[0193] Transmitted content: remote control commands (joystick data, mode switching commands, emergency stop commands, etc.) and status query commands;

[0194] Transport protocol: A custom lightweight protocol based on UDP or TCP;

[0195] Bandwidth requirement: Approximately 10–50 kbps (extremely low).

[0196] (III) Channel Multiplexing Strategy

[0197] The downlink transmission link for video data and the uplink transmission link for control commands are carried in the same wireless channel using time-division multiplexing. Specifically:

[0198] The data transmission module 6 divides the wireless communication time into multiple time slots, with most time slots allocated to downlink transmission of video data (downlink time slots) and a small portion allocated to uplink transmission of control commands (uplink time slots) and downlink transmission of status feedback.

[0199] Uplink time slots appear at a high frequency (e.g., once every 10 milliseconds) to ensure that the transmission delay of remote control commands does not exceed 20 milliseconds, meeting the real-time requirements of remote control. When no drive control commands need to be transmitted, the uplink time slots are automatically given over to video data transmission to improve channel utilization.

[0200] (iv) Communication status monitoring and fault handling

[0201] Data transmission module 6 also has a communication status monitoring function:

[0202] The system periodically (e.g., once per second) sends heartbeat packets to the shore-based monitoring platform and the remote control handle to confirm that the communication link is normal. When a communication link interruption is detected (no response to three consecutive heartbeat packets), the data transmission module 6 sends a communication interruption signal to the core computing power control board 4. After receiving the communication interruption signal, the core computing power control board 4 automatically executes safety protection actions: reducing propulsion speed, controlling the surface vessel to slowly move straight along the current course, or stopping the vessel and waiting for communication to be restored. When the communication link is restored, the core computing power control board 4 automatically re-establishes the data connection with the shore-based monitoring platform and the remote control handle, and restores normal navigation control.

[0203] Example 9 is basically the same as Example 1, except that the multi-chamber sealing structure of the surface vessel body 1 and the waterproof installation design of the multiple cameras 3 are specified in detail.

[0204] (I) Multi-compartment sealed structure of surface vessel body

[0205] The surface vessel hull 1 is integrally molded from high-strength engineering plastics or fiberglass and has multiple independent, sealed compartments. Specifically, the internal space of the surface vessel hull 1 is divided into the following compartments by watertight bulkheads:

[0206] Equipment Cabin: Located in the central region of the surface vessel, this is the largest sealed compartment, housing the core computing power control board 4, data transmission module 6, battery pack, and power management module. The equipment compartment has a removable watertight cover at the top, sealed to the hull with a rubber sealing ring. The cover employs a quick-locking mechanism for easy daily maintenance and repair. The bottom of the equipment compartment has a water-absorbing layer and drainage holes to prevent the accumulation of small amounts of condensate.

[0207] Buoyancy chambers: Distributed at the bow, stern, and sides of the surface vessel, these are fully sealed hollow chambers filled with closed-cell foam material, providing the primary reserve buoyancy. The volume of the buoyancy chambers is calculated based on the surface vessel's full-load displacement and expected load, ensuring that the surface vessel maintains upright stability and floating attitude even if any chamber is flooded.

[0208] Battery compartment: An independent, sealed compartment for housing the power battery pack. The battery compartment is isolated from the equipment compartment by a watertight partition to prevent battery leakage or thermal runaway from affecting core control equipment. The battery compartment is equipped with a waterproof and breathable valve to balance the air pressure inside and outside the compartment.

[0209] Each sealed compartment is isolated from the others by watertight bulkheads, which are at least 3mm thick. The connection between the bulkheads and the hull is sealed using a combination of double-sided sealant and mechanical fastening. In the event of accidental water ingress into any compartment, the remaining compartments still provide sufficient buoyancy to maintain the surface vessel's floating and attitude stability, ensuring that the surface vessel does not sink or capsize.

[0210] (II) Camera Installation and Waterproofing Design

[0211] Each camera 3 is installed at a predetermined position on the surface vessel body 1, and the specific installation method is as follows:

[0212] Bow camera: Mounted on a dedicated camera bracket on the bow deck of the surface vessel, with the bracket height 200mm to 400mm above the deck surface to ensure that the camera's field of view is not obstructed by the bow bulwark. The camera is isolated from the bracket using shock-absorbing balls or rubber pads to reduce the impact of navigation vibrations on image quality.

[0213] Side-mounted cameras: These are installed on the top of the bulwark or on the outer brackets on the port and starboard sides of the surface vessel, respectively. The installation height is kept at the same level as the bow camera to ensure consistent field of view in all directions.

[0214] Each camera is equipped with an independent waterproof housing with a waterproof protection rating of at least IP67 (completely dustproof and capable of being immersed in 1 meter of water for 30 minutes without damage). The waterproof housing is made of aluminum alloy or engineering plastic, and the front of the housing has an optical glass window. The surface of the window is coated with an anti-reflective film and a waterproof and anti-fouling coating to prevent water droplets from adhering and affecting image formation. Thermally conductive silicone is filled between the camera and the waterproof housing, allowing the heat generated by the camera to be conducted to the external environment through the housing, preventing the internal temperature from becoming too high.

[0215] The camera's power and signal cables are connected to the internal wiring harness via waterproof aviation connectors, which have a waterproof rating of at least IP68. The wiring harness is fitted with a waterproof sealing sleeve where it enters the equipment compartment to ensure watertightness.

[0216] (III) Heat dissipation design

[0217] The equipment compartment is equipped with a passive cooling structure: the core computing power control board 4 is tightly attached to an aluminum alloy heat sink via thermal grease, and the heat sink is then connected to the aluminum alloy base plate of the equipment compartment via a thermal pad, transferring the heat generated by the core computing power control board 4 during operation to the bottom plate of the hull, where it is cooled by water flow. The inner walls of the equipment compartment are covered with a heat insulation layer to reduce the impact of external high-temperature radiation on the internal temperature.

[0218] Example 10, as follows Figure 2 As shown, this embodiment describes the complete workflow of the surface vessel for all-area search and rescue according to the present invention, and illustrates the application of the present invention in combination with typical surface rescue scenarios.

[0219] Scenario Description: A person falls into the water in a certain body of water. Rescue personnel arrive at the scene with a search and rescue surface boat covering the entire area, deploy the boat to the water, and activate the system to carry out the search and rescue.

[0220] Phase 1: System Deployment and Startup.

[0221] Rescuers deployed the surface vessel 1 onto the water, ensuring its stability. The power switch was turned on, and the core computing power control board 4 powered on and began system initialization: loading the target detection model weight file into the KPU memory, initializing the camera driver and PWM output interface, and establishing a Wi-Fi communication link with the shore-based monitoring platform 5 and the remote control handle 7. A three-screen real-time view appeared on the display terminal of the shore-based monitoring platform 5, showing the perspective images from the three cameras. The status indicator light on the remote control handle 7 illuminated, indicating that it was currently in manual remote control mode (default mode).

[0222] Phase Two: Manually remotely control the vehicle to the incident area.

[0223] The operator uses the joystick on the remote control handle 7 to control the surface vessel 1 to move from the deployment point to the incident area. In manual remote control mode, the core computing power control board 4 maps the joystick displacement to propulsion speed and steering angle commands, driving the drive mechanism 2 to execute them. Simultaneously, the core computing power control board 4 continuously performs target detection and overlays the detection results onto the screen of the shore-based monitoring platform 5 for the operator's reference. Based on the real-time image displayed on the shore-based monitoring platform and GPS positioning information, the operator steers the surface vessel to the incident area.

[0224] Phase 3: Autonomous Search and Target Discovery.

[0225] When operators spot a suspected target via video feed, they can manually steer the surface boat toward the target location for confirmation and rescue. The manual remote control mode is suitable for scenarios requiring flexible handling of complex situations or where operators lack confidence in autonomous control. Operators can also switch the system to autonomous approach mode by toggling the mode switch on the remote control handle 7. The core computing power control board 4 begins autonomous cruising along a preset search path (such as a bow-shaped scanning path), while simultaneously performing real-time target detection on the three video streams. When a valid target (confidence score ≥ 0.5) suspected to be a person in the water is detected in the camera feed of a certain channel, the core computing power control board 4 automatically switches to target tracking mode and marks the target location with a colored detection box on the corresponding channel screen of the shore-based monitoring platform, while simultaneously displaying the "person" label and confidence score.

[0226] Phase Four: Autonomous Approach and Multi-Perspective Locking.

[0227] The core computing power control board 4 calculates the target's azimuth angle φ and distance d relative to the surface vessel through a multi-view target fusion positioning module. If the target appears in the field of view of two or three cameras simultaneously, the system uses triangulation to perform fusion positioning, improving positioning accuracy. Based on the target's azimuth and distance, the core computing power control board 4 calculates the optimal deflection angle δ and thruster speed n, generating PWM control commands to drive the drive mechanism 2, enabling the surface vessel to automatically navigate towards the target. During the autonomous approach, the shore-based monitoring platform 5 displays the target's azimuth angle, distance, and approach trajectory in real time, allowing operators to monitor the entire search and rescue progress.

[0228] Phase 5: Target Approach and Rescue Operations.

[0229] When the distance between the surface vessel 1 and the target decreases to within a preset rescue distance threshold (e.g., 2 meters), the core computing power control board 4 controls the drive mechanism 2 to stop propulsion, causing the surface vessel to stop near the target. The shore-based monitoring platform 5 displays a notification message, informing the operator that the target has approached the rescue distance. After confirming the target through the screen on the shore-based monitoring platform, the operator can take one of the following two steps to carry out the rescue:

[0230] Method 1 (Automatic Rescue): The operator can trigger the automatic release of the rescue equipment (such as the ejection of an automatic inflatable lifebuoy or rescue float) by pressing the confirmation button on the remote control handle 7 or the "confirm" button on the shore-based monitoring platform.

[0231] Method 2 (Manual Rescue): The operator switches the system back to manual remote control mode and uses the remote control handle to precisely control the position and attitude of the surface boat, cooperating with rescue personnel to complete salvage or towing operations.

[0232] Phase Six: Task Completion and Recovery

[0233] After the rescue mission was completed, the operators used a remote control to steer the surface vessel back to the deployment point for retrieval. The power was then switched off, completing the entire search and rescue mission.

[0234] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A surface vessel for all-terrain search and rescue, characterized in that, include: A surface vessel body (1). A drive mechanism (2) is located at the stern of the surface vessel body (1). The controlled end of the drive mechanism (2) is connected to the control output end of the core computing power control board (4) to receive drive control commands from the core computing power control board (4) and perform propulsion power output and navigation direction turning operations according to the drive control commands. Multiple cameras (3) are arranged in a circumferentially around the body of the surface vessel (1). The field of view of each camera (3) together form a field of view covering the area around the body of the surface vessel (1). Each camera (3) is used to synchronously acquire images of the water surface area around the body of the surface vessel (1) and transmit the acquired multi-channel video streams to the core computing power control board (4). A core computing power control board (4) is mounted inside the surface vessel body (1). The input end of the core computing power control board (4) is connected to the output end of each of the cameras (3). The core computing power control board (4) is used to receive multiple video streams from multiple cameras (3). The core computing power control board (4) is equipped with an automatic detection and recognition driving algorithm. The automatic detection and recognition driving algorithm processes multiple video streams in parallel to detect and identify rescue targets in the omnidirectional area. It generates driving control commands based on the orientation information of the identified rescue targets relative to the surface vessel body (1) and outputs the driving control commands to the driving mechanism (2). A data transmission module (6) is mounted on the surface vessel body (1) and electrically connected to the core computing power control board (4) to establish a wireless data communication link between the core computing power control board (4) and the shore-based monitoring platform (5) and the remote control handle (7); A shore-based monitoring platform (5) establishes a wireless communication link with a data transmission module (6) via HDMI to WIFI. The shore-based monitoring platform (5) is used to receive and display image data from multiple cameras (3) forwarded by the data transmission module (6) in real time. A remote control handle (7) is wirelessly connected to the core computing power control board (4) via a data transmission module (6) to send remote control commands to the core computing power control board (4) and receive execution status information fed back by the core computing power control board (4); The core computing power control board (4) has an autonomous approach mode and a manual remote control mode. In the autonomous approach mode, the core computing power control board (4) autonomously generates drive control commands based on the recognition results of the automatic detection and recognition drive algorithm. In the manual remote control mode, the core computing power control board (4) controls the drive mechanism (2) in response to the remote control commands.

2. The all-terrain search and rescue surface vessel according to claim 1, characterized in that: Multiple cameras (3) are arranged in a distributed array. The field of view of each camera (3) is configured as follows: the field of view of adjacent cameras overlaps at least partially, and the field of view of all cameras is stitched together to form a panoramic field of view covering a 360° range around the body of the boat on the water surface; the horizontal field of view of each camera (3) is not less than 120°.

3. The all-terrain search and rescue surface vessel according to claim 1, characterized in that: The core computing power control board (4) is equipped with a hardware clock synchronization unit. The hardware clock synchronization unit is used to add hardware timestamps to the image frames of each channel when the video streams collected by each camera are input into the core computing power control board. The core computing power control board performs frame-level synchronization and alignment of multiple video streams according to the hardware timestamps of each channel.

4. The all-terrain search and rescue surface vessel according to claim 1, characterized in that: The automatic detection and recognition driving algorithm has a built-in target detection model based on deep learning neural network. The target detection model runs on the neural network processing unit built into the core computing power control board (4) and performs parallel inference operations on multi-channel video streams. The target detection model is implemented with a lightweight network structure and outputs the bounding box coordinates, category confidence score and category label of the target for each frame of input image. The core computing power control board is set with a confidence threshold. Only when the category confidence score is not lower than the confidence threshold is the target determined to be a valid rescue target.

5. The all-terrain search and rescue surface vessel according to claim 4, characterized in that: The automatic detection and recognition driving algorithm also includes a azimuth calculation function. When the target detection model detects the same rescue target in multiple video frames at the same time, the core computing power control board (4) extracts the center coordinates of the bounding box of the target in each camera frame, and combines the pre-calibrated parameters of each camera to solve the azimuth angle φ and distance d of the rescue target relative to the surface vessel body (1) by triangulation or perspective projection method. The core computing power control board (4) calculates the deflection angle δ and the rotation speed n of the drive mechanism according to the deviation Δφ between the current bow direction and the azimuth angle φ, and generates a corresponding PWM control signal to output to the drive mechanism to drive the surface vessel body to sail toward the target.

6. The all-terrain search and rescue surface vessel according to claim 1, characterized in that: The shore-based monitoring platform (5) includes a display terminal and a WIFI receiving module; the core computing power control board (4) combines multiple video streams collected by multiple cameras (3) into a single multi-screen composite video, which is output to the data transmission module (6) via the HDMI interface of the core computing power control board. The HDMI to WIFI encoding unit in the data transmission module (6) encodes the multi-screen composite video and then wirelessly transmits it to the shore-based monitoring platform via the WIFI protocol; the shore-based monitoring platform (5) displays images from the perspectives of each camera in real time on the display terminal in the form of multi-screen splitting, so that operators can grasp the overall situation of the rescue site; at the same time, after detecting the rescue target, the core computing power control board (4) overlays the target location information onto the screen of the corresponding channel of the shore-based monitoring platform (5) and marks the target distance and azimuth parameters in the screen.

7. The surface vessel for all-area search and rescue according to claim 1, characterized in that: The remote control handle (7) is equipped with a mode switching switch. When the mode switching switch is operated, the remote control handle (7) sends a mode switching command to the core computing power control board (4) through the data transmission module (6). In the autonomous approach mode, the joystick operation of the remote control handle (7) is shielded or used only as an auxiliary input. The core computing power control board (4) autonomously plans the trajectory and controls the drive mechanism to execute according to the recognition result of the automatic detection and recognition drive algorithm. In the manual remote control mode, the joystick displacement of the remote control handle (7) is directly mapped to the propulsion speed command and steering angle command of the drive mechanism (2). The operator makes a remote control decision based on the real-time image displayed by the shore-based monitoring platform (5) and issues drive control commands through the remote control handle.

8. A surface vessel for all-area search and rescue according to claim 7, characterized in that: The core computing power control board (4) is configured to, when running in autonomous approach mode, if no effective rescue target is detected within a preset number of consecutive frames, control the drive mechanism (2) to drive the surface vessel body (1) to cruise along a preset search track; if an effective rescue target is detected during the cruise, automatically switch to target tracking mode and control the surface vessel body to sail towards the target; when the distance between the surface vessel body (1) and the target decreases to within a preset rescue distance threshold, the core computing power control board (4) controls the drive mechanism (2) to stop advancing and waits for manual confirmation or triggering the release of rescue equipment.

9. A surface vessel for all-area search and rescue according to claim 1, characterized in that: The data transmission module (6) is used to establish a two-way wireless data communication link between the core computing power control board, the shore-based monitoring platform and the remote control handle. The operating frequency of the data transmission module is 2.4 GHz or 5.8 GHz, and the communication distance is not less than 500 meters. The data transmission module simultaneously carries the downlink transmission link of image data and the uplink transmission link of control commands, so that image transmission and remote control command transmission are carried in the same wireless channel through time division multiplexing and / or frequency division multiplexing.

10. A surface vessel for all-area search and rescue according to claim 1, characterized in that: The surface vessel body has multiple sealed compartments, including an equipment compartment, a buoyancy compartment, and a battery compartment; the equipment compartment is a watertight compartment, and the core computing power control board and data transmission module are housed inside the equipment compartment; each of the cameras is respectively installed on the bow and the top of the side bulwarks of the surface vessel body, and each camera has an independent waterproof shell.

Images