A method and system for capturing collision of a beacon based on shore-based video

Abstract

This invention discloses a method and system for capturing collisions of navigation aids based on shore-based video, relating to the field of waterway navigation aid supervision. First, the shore-based video stream is adaptively enhanced with defogging and image stabilization. A dual-stream deep learning algorithm is then used for pixel-level recognition and semantic segmentation of the vessel and navigation aid. Subsequently, a spatiotemporal interaction model is constructed to calculate the relative motion trend and mask overlap index of the two in real time. Frequency domain analysis technology is used to filter out the periodic wave displacement of the navigation aid and extract the non-periodic abrupt displacement caused by the collision. Finally, based on multi-criteria judgment of overlap and abrupt displacement, the system automatically triggers the extraction and preservation of multi-dimensional evidence. This system includes modules for video acquisition, pixel-level recognition, interactive analysis, interference filtering and judgment, and evidence preservation. This invention solves the pain points of significant interference in the marine environment, difficulty in obtaining collision evidence, and high false alarm rates, significantly enhancing the intelligence and rule of law level of navigation aid supervision.

Description

Technical Field

[0001] This invention relates to the field of monitoring waterway navigation aids, and in particular to a method and system for capturing collisions of navigation aids based on shore-based video. Background Technology

[0002] With the rapid development of waterway transportation, navigation aids, as important navigational aids for guiding ships, positioning, and marking obstacles, are of paramount importance in terms of safety. Traditional navigation aid supervision mainly relies on manual inspections or the installation of collision sensors such as accelerometers. However, manual inspections have poor real-time performance and are difficult to obtain visual evidence of the moment of collision; sensor-based solutions are prone to false alarms in harsh sea conditions due to waves and wind interference, and have high maintenance costs.

[0003] Existing conventional navigational aid video monitoring technologies typically rely on simple motion detection or single target detection logic, failing to effectively distinguish between the periodic swaying of navigational aids caused by waves and the non-periodic displacement caused by ship collisions. Furthermore, they suffer from low detection accuracy in complex lighting conditions, water mist interference, and high wind and wave environments, making it difficult to achieve precise and automated collision evidence locking and monitoring. To address these technical issues, there is an urgent need to develop a more mature navigational aid collision capture method and system based on shore-based video. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for capturing navigational aid collisions based on shore-based video, aiming to solve the problems mentioned in the background art.

[0005] This invention is implemented as follows: On one hand, a method for capturing collisions with navigation aids based on shore-based video, the method comprising: Acquire real-time video streams containing target navigation marks captured by shore-based surveillance cameras, and perform adaptive image enhancement processing on the video streams; A dual-stream deep learning algorithm is used to perform multi-target recognition on the processed video stream, extracting pixel-level features of ship targets and navigation marks in the video frame; By combining motion vector analysis, a spatiotemporal interaction model of ship targets and navigational aid targets is constructed to calculate their relative motion trends and real-time overlap index. Frequency domain analysis is used to filter periodic wave interference noise from navigation beacon targets and extract non-periodic abrupt displacement of the navigation beacon targets; Collision events are determined based on the real-time overlap index and non-periodic abrupt displacement, and the automatic capture and storage of multi-dimensional evidence data are triggered.

[0006] As a further aspect of the present invention, the step of acquiring a real-time video stream containing target navigation marks collected by a shore-based monitoring camera and performing adaptive image enhancement processing on the video stream specifically includes: Establish a low-latency streaming media transmission link to extract raw video frames in real time; A dark channel prior algorithm is used to dehaze and restore the perspective of video frames, eliminating interference from sea surface mist. Extract the global brightness average of video frames and apply a dynamic gamma correction algorithm to balance the brightness of images with drastic changes in light and shadow. Electronic image stabilization compensation is performed using a feature point matching algorithm to filter out global pixel shifts caused by vibrations of the shore-based support structure.

[0007] As a further aspect of the present invention, the step of using a dual-stream deep learning algorithm to perform multi-target recognition on the processed video stream and extracting pixel-level features of ship targets and navigational aid targets in the video frame specifically includes: The video frames are input into a preset feature extraction backbone network to obtain multi-scale feature maps. The semantic segmentation branch is used to delineate the edge contours of navigation marks in the video image and generate a navigation mark pixel mask. The target detection branch is used to identify the type and bounding box of the ship target, and to calculate the centroid coordinates of the ship target; By combining camera parameters for monocular ranging, the target scale in the pixel coordinate system is converted into the physical scale in the geographic information system.

[0008] As a further aspect of the present invention, the step of combining motion vector analysis to construct a spatiotemporal interaction model between the ship target and the navigation mark target, and calculating their relative motion trend and real-time overlap index specifically includes: Continuously track the temporal positions of ship targets and navigational aids to obtain their motion trajectory sequences; The Kalman filter algorithm is used to predict the expected coverage area of ​​the ship target in the next frame; The intersection-union ratio of the ship target bounding box and the navigation mark pixel mask on the two-dimensional projection plane is calculated to obtain the real-time overlap index; Calculate the radial relative velocity of the ship approaching the navigation target to assess the kinetic energy hazard level of a potential collision.

[0009] As a further aspect of the present invention, the step of determining collision events based on the real-time overlap index and non-periodic abrupt displacement, and triggering the automatic interception and storage of multi-dimensional forensic data, specifically includes: The trajectory of the center of mass of the navigation beacon is subjected to a fast Fourier transform to analyze its amplitude distribution in the frequency domain. Identify and filter low-frequency oscillation components that match wave cycle characteristics; The presence of instantaneous high-frequency impact displacements exceeding a threshold within the monitoring time domain is defined as non-periodic abrupt displacement. When the collision determination takes effect, high-definition video clips before and after the collision, collision location metadata, and target feature reports are captured simultaneously, and then encrypted, packaged, and uploaded.

[0010] As a further aspect of the present invention, another option is a navigation mark collision detection system based on shore-based video, the system comprising: The video adaptive acquisition module is used to perform video stream acquisition, dehazing, brightness correction, and image stabilization enhancement processing; A pixel-level target recognition module is used to achieve semantic segmentation, feature extraction, and physical scale mapping of ships and navigation marks; The spatiotemporal interaction analysis module is used to construct motion trajectory models and calculate the overlap and relative motion trends between targets; The interference filtering and determination module is used to eliminate wave interference through frequency domain analysis and calculate the abnormal sudden displacement of the navigation mark; The intelligent capture and evidence storage module is used to capture, encapsulate, and push multi-dimensional evidence to the cloud based on the judgment results.

[0011] As a further aspect of the present invention, the pixel-level target recognition module specifically includes: Multi-task backbone neural network units are used to extract deep semantic features of video frames using deep convolutional layers; The mask generation and segmentation unit is used to perform pixel-level classification of navigational targets and output closed edge contour curves. The spatial geometric ranging unit is used to combine the camera intrinsic parameter matrix and the installation height parameter to transform the pixel coordinate system to the geographic coordinate system, thereby realizing real-time quantitative measurement of the physical Euclidean distance between the ship and the navigation mark.

[0012] As a further aspect of the present invention, the spatiotemporal interaction analysis module specifically includes: The multi-target tracking association unit is used to assign globally unique UIDs to multiple dynamic targets appearing in the video frame and maintain their historical motion trajectory chains to solve the ID jump problem caused by wave occlusion. The collision risk prediction unit is used to extrapolate and predict the inertial motion trajectory of a ship using state-space equations, and to assess the probability of overlap between the ship's track and the stationary area of ​​the navigation mark. The mask-level overlap calculation unit is used to generate a real-time overlap index to characterize the degree of contact by calculating the intersection-union ratio of the ship detection frame and the edge mask of the navigation mark.

[0013] This invention provides a method and system for detecting navigational aid collisions based on shore-based video. Through the deep fusion of deep learning semantic segmentation and frequency domain analysis, it significantly improves the accuracy of navigational aid collision monitoring under complex sea conditions. It effectively filters out periodic oscillation interference caused by ocean waves, achieving millisecond-level capture of non-periodic abrupt displacements at the moment of collision. Furthermore, through pixel-mask-level overlap calculation, it completely avoids the risk of false alarms caused by visual occlusion in traditional solutions. Attached Figure Description

[0014] Figure 1 This is the main flowchart of a navigation mark collision detection method based on shore-based video.

[0015] Figure 2 This is a flowchart illustrating a shore-based video-based method for capturing collisions between navigational aids. It combines motion vector analysis to construct a spatiotemporal interaction model between the ship target and the navigational aid target, calculating their relative motion trend and real-time overlap index.

[0016] Figure 3 This is a flowchart illustrating the automatic capture and storage of multi-dimensional evidence data in a navigation mark collision capture method based on shore-based video, which determines collision events based on the real-time overlap index and non-periodic abrupt displacement.

[0017] Figure 4 This is a main structure diagram of a navigation mark collision detection system based on shore-based video. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0020] The present invention provides a method and system for capturing collisions of navigation marks based on shore-based video, which solves the technical problems in the background art.

[0021] like Figure 1 The diagram shown is a main flowchart of a navigational aid collision detection method based on shore-based video, according to an embodiment of the present invention. The method includes: Step S100: Acquire the real-time video stream containing the target navigation mark collected by the shore-based monitoring camera, and perform adaptive image enhancement processing on the video stream; Step S200: Use a dual-stream deep learning algorithm to perform multi-target recognition on the processed video stream and extract pixel-level features of ship targets and navigation marks in the video frame; Step S300: Combining motion vector analysis, construct a spatiotemporal interaction model of the ship target and the navigation target, and calculate the relative motion trend and real-time overlap index between the two. Step S400: Use frequency domain analysis to filter periodic wave interference noise from navigation beacon targets and extract non-periodic abrupt displacement of navigation beacon targets; Step S500: Based on the real-time overlap index and the non-periodic abrupt displacement, determine the collision event and trigger the automatic capture and storage of multi-dimensional evidence data; In this embodiment, the system first uses a shore-based high-definition infrared night vision camera to acquire a real-time video stream covering a specific waterway area using the RTSP protocol. Considering the complex ambient lighting at sea, the system's built-in adaptive image enhancement program adjusts the contrast in real time based on the histogram distribution of the image. Then, a dual-stream deep learning architecture based on a residual network is employed. One convolutional branch identifies the bounding box of the ship, while the other fully convolutional branch obtains the pixel mask of the navigation mark through semantic segmentation, thus accurately locking the pixel boundary between the two amidst wave fluctuations. After obtaining the target features, a spatiotemporal interaction model is constructed using a multi-frame association algorithm. This model not only calculates the Euclidean distance between the ship's centroid and the navigation mark's centroid but also calculates the intersection-over-union (IoU) ratio of the ship's boundary and the navigation mark's pixel mask in real time. To eliminate interference from waves, a frequency domain transformation is performed on the navigation mark's motion sequence to separate the low-frequency reciprocating displacement caused by waves from the high-frequency pulse displacement generated by collisions. When the system detects a physical contact trend between the ship and the navigation mark and the navigation mark shows abnormal non-periodic displacement, the logic adjudication module will immediately send an interrupt signal to the cache management component, retrieve the high-definition video of 15 seconds before and after the collision stored in the circular buffer, and encapsulate it in a structured manner along with the geographical location, timestamp, and target feature ID at that time, and push it to the maritime supervision platform through the private network to complete the entire process from monitoring to evidence storage.

[0022] In a preferred embodiment of the present invention, the step of acquiring a real-time video stream containing target navigation marks collected by a shore-based monitoring camera and performing adaptive image enhancement processing on the video stream specifically includes: Step S101: Establish a low-latency streaming media transmission link and extract raw video frames in real time; Step S102: Use a dark channel prior algorithm to dehaze and restore the perspective of the video frames to eliminate interference from sea surface mist; Step S103: Extract the global brightness mean of the video frame and apply the dynamic gamma correction algorithm to balance the brightness of the scene with drastic changes in light and shadow. Step S104: Use the feature point matching algorithm to perform electronic image stabilization compensation and filter out global pixel shifts caused by vibration of the shore support.

[0023] In this embodiment, a dual buffer is first allocated in memory, and a low-latency streaming link is established through the streaming media scheduling unit to extract the original video frames in real time. When facing common maritime interference such as sea fog or water vapor, the dark channel prior algorithm is invoked to perform global dehazing by calculating the transmittance map of the video frames and combining it with atmospheric light estimation values, restoring details lost due to scattering. Simultaneously, for high-contrast scenes such as sunrise and sunset, the system extracts the global brightness average of each frame. If the average deviates from a preset threshold, the dynamic gamma correction coefficient is automatically adjusted to suppress overexposure and enhance shadow details. The system utilizes the ORB feature point matching algorithm to find stationary shore-based reference objects (such as distant mountains or fixed docks) between consecutive video frames, and calculates the affine transformation matrix to correct global pixel shifts caused by sea winds or base vibrations.

[0024] In a preferred embodiment of the present invention, the step of using a dual-stream deep learning algorithm to perform multi-target recognition on the processed video stream and extracting pixel-level features of ship targets and navigation marks in the video frame specifically includes: Step S201: Input the video frames into a preset feature extraction backbone network to obtain multi-scale feature maps; Step S202: Use the semantic segmentation branch to delineate the edge contours of the navigation beacon targets in the video image and generate a navigation beacon pixel mask; Step S203: Use the target detection branch to identify the type and bounding box of the ship target, and calculate the centroid coordinates of the ship target; Step S204: Combine camera parameters to perform monocular ranging, and convert the target scale in the pixel coordinate system into the physical scale in the geographic information system.

[0025] In this embodiment, the enhanced video frames are input into a pre-trained backbone network, where branching processing is performed at the feature layer level. The target detection branch uses an improved YOLO architecture to perform multi-scale bounding box selection for ship targets, accurately identifying different types such as bulk carriers and container ships in foggy or small target scenarios, and outputting the dynamic trajectory of their centroid in the pixel coordinate system in real time. Simultaneously, the semantic segmentation branch uses a lightweight DeepLabV3+ model to perform single-class pixel-level segmentation of navigation mark targets, generating binary masks accurate to pixel edges, enabling the system to perceive the shape changes of navigation marks in waves. To convert visual information into physical information, a monocular ranging model is introduced, using the camera's focal length, installation pitch angle, installation height, and sea level as reference planes. The pixel coordinates of the bottom center point of the ship's bounding rectangle are mapped to latitude and longitude geographic coordinates using a homography matrix.

[0026] like Figure 2As shown, in a preferred embodiment of the present invention, the step of constructing a spatiotemporal interaction model of the ship target and the navigation mark target by combining motion vector analysis, and calculating the relative motion trend and real-time overlap index of the two specifically includes: Step S301: Continuously track the temporal positions of the ship target and the navigation beacon target to obtain their motion trajectory sequence; Step S302: Predict the expected coverage area of ​​the ship target in the next frame using the Kalman filter algorithm; Step S303: Calculate the intersection-union ratio of the outer bounding box of the ship target and the pixel mask of the navigation mark on the two-dimensional projection plane to obtain the real-time overlap index; Step S304: Calculate the radial relative velocity of the ship target approaching the navigation beacon target and assess the kinetic energy hazard level of the potential collision.

[0027] It should be understood that the system assigns a globally unique UID to each ship and navigation mark appearing in the video and maintains its motion chain through a multi-target tracking algorithm. In each frame processing, the Kalman filter algorithm is used to extrapolate the expected position of the target in the current frame based on the motion vectors of the previous five frames, forming a prediction coverage area. To determine the possibility of a collision, a mask-level intersection-union (IUU) calculation is introduced: the overlap between the predicted bounding box of the ship and the real-time semantic segmentation mask of the navigation mark is analyzed. If the overlap area between the two in the 2D image exceeds 5% of the total area, and the radial relative velocity of the ship approaching the navigation mark shows a continuous increasing trend, the system determines it to be a high-risk interaction state. At this time, the system further combines the physical scale obtained from monocular ranging to calculate the shortest geometric distance between the edge of the ship and the edge of the navigation mark in geographic space. By constructing this interaction model based on dual verification of pixel overlap and physical distance, the system can effectively distinguish between two scenarios: the ship occluding the navigation mark due to viewing angle and the ship actually touching the navigation mark, greatly improving the accuracy of collision recognition.

[0028] like Figure 3 As shown, in a preferred embodiment of the present invention, the step of determining collision events based on the real-time overlap index and non-periodic abrupt displacement, and triggering the automatic capture and storage of multi-dimensional forensic data, specifically includes: Step S501: Perform a fast Fourier transform on the trajectory of the center of mass of the navigation beacon target and analyze its amplitude distribution in the frequency domain; Step S502: Identify and filter low-frequency oscillation components that match the wave cycle characteristics; Step S503: Monitor whether there is an instantaneous high-frequency impact displacement exceeding the threshold in the time domain, and define it as a non-periodic abrupt displacement. Step S504: When the collision determination takes effect, simultaneously capture high-definition video clips before and after the collision, collision location metadata, and target feature reports, and encrypt and encapsulate them before uploading.

[0029] In this embodiment, the motion trajectory of the centroid of the buoy pixel mask is recorded in real time, forming a time-series signal. Considering that the buoy will oscillate quasi-periodically due to the influence of waves on the sea surface, a Fast Fourier Transform is used to convert the motion signal from the time domain to the frequency domain. By analyzing the spectrum, the low-frequency peak characteristics corresponding to the wave cycle can be identified and filtered out using a band-stop filter, thereby eliminating conventional displacement noise caused by the natural environment. After filtering out background noise, the time-domain performance of the residual signal is monitored in real time. Once a transient, non-periodic, pulse-like high-frequency impact displacement exceeding the statistical standard deviation is detected, it is determined to be a mechanical response of a physical collision. At this time, the collision decision module immediately triggers the evidence storage logic, extracts the original video stream containing 5 seconds before to 25 seconds after the moment of collision from the memory circular queue, and overlays metadata such as the current time, ship characteristics, and trajectory coordinates to generate an unalterable H.265 encrypted video file. This file is marked as the highest priority and uploaded to the cloud server via network slicing technology, and an audible and visual alarm is issued on the monitoring terminal.

[0030] like Figure 4 As shown, in another preferred embodiment of the present invention, a navigation mark collision detection system based on shore-based video is provided, the system comprising: The video adaptive acquisition module 100 is used to perform video stream acquisition, dehazing, brightness correction and anti-shake enhancement processing; The pixel-level target recognition module 200 is used to realize semantic segmentation, feature extraction and physical scale mapping of ships and navigation marks; The spatiotemporal interaction analysis module 300 is used to construct motion trajectory models and calculate the overlap and relative motion trends between targets. The interference filtering and determination module 400 is used to eliminate wave interference through frequency domain analysis and calculate the abnormal sudden displacement of the navigation mark. The intelligent capture and evidence storage module 500 is used to capture, encapsulate, and push multi-dimensional evidence to the cloud based on the judgment results.

[0031] In this embodiment, the video adaptive acquisition module 100 acts as a data engine, responsible for maintaining long-term connections with multiple front-end cameras; the pixel-level target recognition module 20 acts as the perception brain, utilizing the GPU hardware acceleration unit to process deep learning inference tasks in parallel; the spatiotemporal interaction analysis module 300 acts as the logic hub, calculating the dynamic relationships between targets in real time; the interference filtering judgment module 400 acts as a decision filter, verifying the authenticity of collisions through computationally intensive spectrum analysis; and the intelligent capture and evidence storage module 500 acts as the execution terminal, managing video slices in the solid-state drive to ensure stable transmission of high-definition evidence under limited bandwidth.

[0032] In another preferred embodiment of the present invention, the pixel-level target recognition module 200 specifically includes: The multi-task backbone neural network unit 201 is used to extract deep semantic features of video frames using deep convolutional layers. The mask generation and segmentation unit 202 is used to perform pixel-level classification of navigation target and output closed edge contour curves; The spatial geometric ranging unit 203 is used to combine the camera intrinsic parameter matrix and the installation height parameter to transform the pixel coordinate system to the geographic coordinate system, so as to realize the real-time quantitative measurement of the physical Euclidean distance between the ship and the navigation mark.

[0033] In this embodiment, the multi-task backbone neural network unit 201 employs depthwise separable convolution technology, which greatly improves computational efficiency while ensuring recognition accuracy, enabling real-time simulation at 30fps on embedded edge devices. The mask generation and segmentation unit 202 fuses features of different depths through a feature pyramid network, enabling it not only to identify large merchant ships but also to perform pixel-level segmentation of slender, long-distance navigation marks. This allows the system to capture the minute deformation or tilt of the navigation mark at the moment of impact. The spatial geometry ranging unit 203 incorporates an offline calibration database, injecting initialization parameters for the installation height, pitch angle, and geographic coordinates of each camera. During operation, this unit uses the camera calibration model to calculate the intersection of the segmented bottom pixel coordinates of the navigation mark with the geographic projection surface, outputting accurate latitude and longitude.

[0034] In another preferred embodiment of the present invention, the spatiotemporal interaction analysis module 300 specifically includes: The multi-target tracking association unit 301 is used to assign globally unique UIDs to multiple dynamic targets appearing in the video frame and maintain their historical motion trajectory chains to solve the ID jump problem caused by wave occlusion. The collision risk prediction unit 302 is used to extrapolate and predict the inertial motion trajectory of the ship using state-space equations, and to assess the probability of overlap between the ship's track and the stationary area of ​​the navigation mark. The mask-level overlap calculation unit 303 is used to generate a real-time overlap index to characterize the degree of contact by calculating the intersection-union ratio of the ship detection frame and the edge mask of the navigation mark.

[0035] In this embodiment, the multi-target tracking association unit 301 introduces an appearance feature re-identification algorithm. When a ship passes a navigation mark and the navigation mark is temporarily obscured due to overlapping views, the system locks the navigation mark's UID and quickly restores the association after the obscuration ends, avoiding monitoring blind spots caused by trajectory interruptions. The collision risk prediction unit 302 calculates the shortest spatiotemporal distance between the ship's track tangent and the navigation mark protection zone to estimate the collision time in advance. The mask-level overlap calculation unit 303 uses GPU-accelerated logic operations to quickly solve for the intersection area of ​​two polygonal regions. Compared to the traditional two-frame collision alarm scheme based on rectangular box detection, this can accurately distinguish whether the visual overlap caused by the ship passing behind the navigation mark is a real collision occurring at the same physical level.

[0036] The above embodiments of the present invention provide a method and system for detecting navigational aid collisions based on shore-based video. Through deep fusion of deep learning semantic segmentation and frequency domain analysis, the accuracy of navigational aid collision monitoring under complex sea conditions is significantly improved. It effectively filters out periodic oscillation interference caused by ocean waves, achieving millisecond-level capture of non-periodic abrupt displacement at the moment of collision; and through pixel-mask-level overlap calculation, it completely avoids the risk of false alarms caused by visual occlusion in traditional solutions.

[0037] In order for the above methods and systems to operate smoothly, the system may include more or fewer components than those described above, or combine certain components, or different components, in addition to the various modules mentioned above. For example, it may include input / output devices, network access devices, buses, processors, and memory.

[0038] The processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (OPGs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the system, connecting various parts via various interfaces and lines.

[0039] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0040] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0041] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for capturing collisions of navigation marks based on shore-based video, characterized in that, The method includes: Acquire real-time video streams containing target navigation marks captured by shore-based surveillance cameras, and perform adaptive image enhancement processing on the video streams; A dual-stream deep learning algorithm is used to perform multi-target recognition on the processed video stream, extracting pixel-level features of ship targets and navigation marks in the video frame; By combining motion vector analysis, a spatiotemporal interaction model of ship targets and navigational aid targets is constructed to calculate their relative motion trends and real-time overlap index. Frequency domain analysis is used to filter periodic wave interference noise from navigation beacon targets and extract non-periodic abrupt displacement of the navigation beacon targets; Collision events are determined based on the real-time overlap index and non-periodic abrupt displacement, and the automatic capture and storage of multi-dimensional evidence data are triggered.

2. The navigation mark collision detection method based on shore-based video according to claim 1, characterized in that, The process of acquiring real-time video streams containing target navigation marks captured by shore-based surveillance cameras and performing adaptive image enhancement processing on the video stream specifically includes: Establish a low-latency streaming media transmission link to extract raw video frames in real time; A dark channel prior algorithm is used to dehaze and restore the perspective of video frames, eliminating interference from sea surface mist. Extract the global brightness average of video frames and apply a dynamic gamma correction algorithm to balance the brightness of images with drastic changes in light and shadow. Electronic image stabilization compensation is performed using a feature point matching algorithm to filter out global pixel shifts caused by vibrations of the shore-based support structure.

3. The navigation mark collision detection method based on shore-based video according to claim 1, characterized in that, The step of using a dual-stream deep learning algorithm to perform multi-target recognition on the processed video stream and extracting pixel-level features of ship targets and navigational aids in the video footage specifically includes: The video frames are input into a preset feature extraction backbone network to obtain multi-scale feature maps. The semantic segmentation branch is used to delineate the edge contours of navigation marks in the video image and generate a navigation mark pixel mask. The target detection branch is used to identify the type and bounding box of the ship target, and to calculate the centroid coordinates of the ship target; By combining camera parameters for monocular ranging, the target scale in the pixel coordinate system is converted into the physical scale in the geographic information system.

4. The navigation mark collision detection method based on shore-based video according to claim 1, characterized in that, The process of constructing a spatiotemporal interaction model between the ship target and the navigation mark target by combining motion vector analysis, and calculating their relative motion trend and real-time overlap index, specifically includes: Continuously track the temporal positions of ship targets and navigational aids to obtain their motion trajectory sequences; The Kalman filter algorithm is used to predict the expected coverage area of ​​the ship target in the next frame; The intersection-union ratio of the ship target bounding box and the navigation mark pixel mask on the two-dimensional projection plane is calculated to obtain the real-time overlap index; Calculate the radial relative velocity of the ship approaching the navigation target to assess the kinetic energy hazard level of a potential collision.

5. The navigation mark collision detection method based on shore-based video according to claim 1, characterized in that, The process of determining collision events based on the real-time overlap index and non-periodic abrupt displacement, and triggering the automatic capture and storage of multi-dimensional evidence data, specifically includes: The trajectory of the center of mass of the navigation beacon is subjected to a fast Fourier transform to analyze its amplitude distribution in the frequency domain. Identify and filter low-frequency oscillation components that match wave cycle characteristics; The presence of instantaneous high-frequency impact displacements exceeding a threshold within the monitoring time domain is defined as non-periodic abrupt displacement. When the collision determination takes effect, high-definition video clips before and after the collision, collision location metadata, and target feature reports are captured simultaneously, and then encrypted, packaged, and uploaded.

6. A navigational aid collision detection system based on shore-based video, characterized in that, The system employs the shore-based video-based navigational aid collision detection method as described in any one of claims 1-5, wherein the system comprises: The video adaptive acquisition module is used to perform video stream acquisition, dehazing, brightness correction, and image stabilization enhancement processing; A pixel-level target recognition module is used to achieve semantic segmentation, feature extraction, and physical scale mapping of ships and navigation marks; The spatiotemporal interaction analysis module is used to construct motion trajectory models and calculate the overlap and relative motion trends between targets; The interference filtering and determination module is used to eliminate wave interference through frequency domain analysis and calculate the abnormal sudden displacement of the navigation mark; The intelligent capture and evidence storage module is used to capture, encapsulate, and push multi-dimensional evidence to the cloud based on the judgment results.

7. The navigation mark collision detection system based on shore-based video according to claim 6, characterized in that, The pixel-level target recognition module specifically includes: Multi-task backbone neural network units are used to extract deep semantic features of video frames using deep convolutional layers; The mask generation and segmentation unit is used to perform pixel-level classification of navigational targets and output closed edge contour curves. The spatial geometric ranging unit is used to combine the camera intrinsic parameter matrix and the installation height parameter to transform the pixel coordinate system to the geographic coordinate system, thereby realizing real-time quantitative measurement of the physical Euclidean distance between the ship and the navigation mark.

8. The navigation mark collision detection system based on shore-based video according to claim 6, characterized in that, The spatiotemporal interaction analysis module specifically includes: The multi-target tracking association unit is used to assign globally unique UIDs to multiple dynamic targets appearing in the video frame and maintain their historical motion trajectory chains to solve the ID jump problem caused by wave occlusion. The collision risk prediction unit is used to extrapolate and predict the inertial motion trajectory of a ship using state-space equations, and to assess the probability of overlap between the ship's track and the stationary area of ​​the navigation mark. The mask-level overlap calculation unit is used to generate a real-time overlap index to characterize the degree of contact by calculating the intersection-union ratio of the ship detection frame and the edge mask of the navigation mark.

Images