A method, recording medium and system for detecting a moving target on a water surface

Abstract

The present application belongs to the technical field of image data processing, and particularly relates to a detection method of a water surface moving target, comprising: filtering, aligning, voxelizing and constructing a global voxel map for target original point cloud data obtained by a laser radar and a synchronous inertial measurement unit, assigning different dimensions of observation attributes to each voxel, obtaining an initial selected dynamic voxel set from a space-time perspective, extracting voxel features in the set, constructing a point cloud feature descriptor, further accurately detecting dynamic voxels by using geometric feature differences, merging adjacent dynamic voxels, and obtaining a moving target. The method can correlate voxel space-time features and geometric features under the condition that environmental parameters are unknown and point clouds are sparse, and takes into account detection accuracy and real-time performance. The method is suitable for water surface moving target detection. The present application also provides a non-transient readable recording medium storing a program of the method and a system comprising the medium, and the program can be called by a processing circuit to execute the above method.

Description

Technical Field

[0001] This invention belongs to the field of image data processing technology and discloses a method, recording medium and system for detecting moving targets on the water surface. Background Technology

[0002] Unmanned surface vessels (USVs), as core equipment of intelligent waterway transportation systems, are widely used in tasks such as water surface mapping, emergency search and rescue, and waterway patrol. Moving target detection, a crucial component of USV environmental perception, requires providing real-time and reliable dynamic target information to ensure safe navigation in complex water environments. However, the unique characteristics of water surface scenarios present numerous challenges to moving target detection:

[0003] Poor sensor data quality: Overexposure of light on the water surface, foggy weather and other environments can easily lead to sparse point clouds of lidar, and even generate false data, which can interfere with the extraction of target features;

[0004] The target motion characteristics are complex: Surface targets (such as buoys and ships) will sway irregularly due to the influence of waves, causing the point cloud position to drift and increasing the difficulty of distinguishing between dynamic and static targets;

[0005] The target size varies greatly: the size of water surface targets ranges from a few meters (small boats) to hundreds of meters (large ships), and traditional methods are difficult to balance the detection accuracy of targets of different sizes.

[0006] The conflict between real-time performance and computing power: Existing neural network-based lidar moving target detection methods have high computational overhead, while small USVs with limited carrying capacity are difficult to equip with high-performance computing devices, thus failing to meet the real-time detection requirements.

[0007] Existing LiDAR moving target detection methods include some that combine voxelization with map building, but these methods suffer from weak point cloud noise suppression and poor adaptability to sparse point clouds over water. Other lightweight methods reduce computational requirements, but they do not adequately consider scenarios such as water surface target movement and long-distance sparse point clouds, resulting in insufficient detection robustness. Summary of the Invention

[0008] To address the above problems, this invention provides a method for detecting moving targets on the water surface, comprising the following steps:

[0009] S1. The raw point cloud data of the target obtained by the lidar and synchronous inertial measurement unit is filtered and then preprocessed with temporal and spatial alignment.

[0010] S2. Divide the preprocessed point cloud into voxels, construct a real-time updated global voxel map, and assign different dimensions of observation attributes to each voxel to obtain a preliminary dynamic voxel set from a spatiotemporal perspective.

[0011] S3. Within the initial set of dynamic voxels, voxel features are obtained using the covariance matrix. Local and global features and surface curvature information are fused to construct a point cloud feature descriptor, which characterizes the geometric feature differences of voxels, distinguishes between dynamic and non-dynamic voxels, and removes non-dynamic voxels from the set of dynamic voxels to obtain the set of corrected dynamic voxels.

[0012] S4. Using the corrected dynamic voxels as seeds, perform region growth, merge adjacent corrected dynamic voxels, and obtain the moving target.

[0013] Preferably, time alignment is achieved by mapping all data to the same time axis through a timestamp synchronization module, thus realizing the temporal alignment of multi-sensor data; spatial alignment is achieved by accurately mapping the lidar point cloud of the previous frame to the reference coordinate system of the next frame through pose transformation; filtering includes: height threshold filtering to remove reflection points located below or adjacent to the water surface and neighborhood density filtering to filter reflection points with neighborhood density less than the threshold.

[0014] Preferably, obtaining the initial dynamic voxel set from a spatiotemporal perspective includes the following steps: For each voxel, it can be represented by multiple observation attribute definition states: Each attribute is defined as follows:

[0015] D i To truncate the cumulative value of the TSDF (Signal Distance Function), D i =0 indicates that there is no accumulated information in the voxels;

[0016] W i To truncate the cumulative observation weights of the TSDF (Signal Distance Function), W i =0, meaning the voxel was not observed;

[0017] The most recently occupied timestamp for the current voxel; This indicates that the space was not initially occupied.

[0018] For continuous occupation duration; This indicates that the initial occupancy was not contiguous;

[0019] f is the voxel free state discrimination value, f=0 indicates non-free, f=1 indicates free. Initially, it is non-free by default. If the current voxel satisfies the condition that the voxel's TSDF value is greater than the preset threshold, then f is allowed to determine the free state. , or there may be a point cloud in the current frame. If it falls into that voxel, then it is updated. If a voxel is not free, it is considered a non-free voxel; otherwise, it is considered a free voxel. Set an N value, and for the current non-free voxel, if it was free in the previous N frames, it is determined to be a dynamic voxel. Then, all the voxels determined to be dynamic are collected into a set of dynamic voxels.

[0020] Preferably, the point cloud feature descriptor also includes point curvature calculation, which supplements the feature dimension in distorted scenes by utilizing the stable representation characteristics of curvature on the target geometry.

[0021] Preferably, the geometric feature differences of voxels, characterized by the point cloud covariance matrix, are reflected by distance in Riemann space.

[0022] Preferably, during the region growing process, the number of voxels in the dynamic voxel cluster obtained by the region growing algorithm is less than the minimum number of voxels in a single cluster. If a cluster is identified as noise, it is removed from the dynamic voxel set.

[0023] Another aspect of the present invention is to provide a non-transient readable recording medium for storing one or more programs containing multiple instructions, which, when executed, cause a processing circuit to perform the aforementioned method for detecting a moving target on the water surface.

[0024] Another aspect of the present invention provides a detection system for moving targets on the water surface, including a processing circuit and a memory electrically coupled thereto. The memory is configured to store at least one program, the program containing multiple instructions. The processing circuit runs the program and can execute the above-described method for detecting moving targets on the water surface.

[0025] Compared to existing technologies, the present invention provides a method, recording medium, and system for detecting moving targets on the water surface. This method constructs a dynamically updated map using voxelization, assigns spatiotemporal attributes to voxels for initial dynamic voxel selection, compresses the amount of data required for later dynamic target detection, and improves detection efficiency. It also designs a feature descriptor that integrates local-global features and curvature information, improving feature matching accuracy under sparse point clouds and reducing misjudgments caused by target movement. Furthermore, it introduces a neighborhood dynamic weight and distance-sensitive parameter module to suppress point cloud drift and noise interference, solving the accuracy problem caused by sparse long-distance point clouds, ultimately achieving real-time and accurate detection of moving targets in water surface scenarios. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the process of detecting a moving target on the water surface in an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of feature descriptor construction in an embodiment of the present invention;

[0028] Figure 3 This is a diagram showing the effect of moving target detection in a general scenario;

[0029] Figure 4 This is a diagram showing the detection effect of the present invention in a water surface scenario;

[0030] Figure 5 This is a performance comparison chart of the present invention and various other methods in a general scenario;

[0031] Figure 6 This is a performance comparison chart of the present invention and various other methods in a water surface scenario. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without innovative effort are within the scope of protection of the present invention.

[0033] Please refer to the accompanying drawings. An embodiment of the present invention provides a method for detecting moving targets on the water surface, comprising the following steps:

[0034] S1: Based on the motion characteristics of unmanned surface vessels and the characteristics of lidar data, the input raw point cloud and synchronous IMU data are preprocessed to provide a high-quality data foundation for subsequent voxelization and map construction.

[0035] S2: Divide the preprocessed point cloud into voxels, construct a real-time updated global voxel map, and assign multi-dimensional attributes to each voxel to preliminarily identify dynamic voxels from a spatiotemporal perspective.

[0036] S3: Based on voxel features, local and global features and surface curvature information are integrated to construct a robust point cloud feature descriptor. Features are represented by the covariance matrix to achieve effective feature matching under sparse point clouds.

[0037] S4: Combining voxel attributes and feature descriptors, design dynamic voxel classification rules, and further verify the dynamic nature of voxels through feature matching to eliminate misjudgments caused by noise interference.

[0038] S5: Region growth is performed using dynamic voxels as seeds, and adjacent dynamic voxels are merged. The model performance is tested in real water surface scenarios and conventional scenarios to verify the detection accuracy and real-time performance of the method under different distances and levels of interference, ensuring that it meets the application requirements of unmanned surface vessels.

[0039] Furthermore, step S1 includes the following steps:

[0040] S11: Spatiotemporal alignment of multi-source data:

[0041] When detecting moving targets on the water surface, the system often involves the joint operation of multiple sensors. Data alignment and coordinate transformation between the IMU and LiDAR are crucial for ensuring detection accuracy. Specifically, because the IMU and LiDAR have different sampling frequencies and trigger times, directly using their raw data will lead to temporal inconsistencies, affecting the accurate representation of target motion. Therefore, a timestamp synchronization module is needed to map the data from the IMU and LiDAR onto the same timeline, achieving temporal alignment of the multi-sensor data.

[0042] After timestamp alignment, the coordinates of the two consecutive LiDAR images need to be transformed using the attitude and displacement information provided by the IMU. During system operation, the attitude and displacement information provided by the IMU can be represented as a rotation matrix. With translation vector The coordinate transformation relationship between the two point clouds is as follows:

[0043] (1)

[0044] in This represents the point cloud coordinates at time t. Let represent the point cloud coordinates at time t+1. Through the transformation of formula (1), the point cloud of the previous frame of lidar can be accurately mapped to the reference coordinate system of the next frame.

[0045] After the above transformation, point cloud data collected at different times are unified under the same spatial reference frame. The system can accurately analyze the displacement and changes of the target between consecutive frames, thereby improving the stability and robustness of detecting moving targets on the water surface. This method not only ensures the temporal and spatial consistency of sensor data, but also provides a reliable data foundation for subsequent target matching and motion estimation.

[0046] S12: Water surface point cloud data preprocessing:

[0047] When detecting moving targets on the water surface, environmental factors often result in a large number of clutter points generated by water surface reflection in the lidar point cloud. These clutter points have low intensity and irregular spatial distribution; if left untreated, they can interfere with subsequent target detection and recognition. Therefore, a water surface clutter filtering step needs to be added during the data preprocessing stage. This step specifically includes two sub-processes: height threshold filtering and neighborhood density filtering.

[0048] Height threshold filtering: Since most clutter points are distributed near the water surface, height constraints can usually be used to remove invalid points. Let the original point cloud data be:

[0049] (2)

[0050] in Let be the three-dimensional coordinates of the i-th point, and N be the total number of points. A water surface threshold is set based on the sensor installation height and the water surface position. In this invention, point clouds representing the water surface that are below this height are filtered out. The lidar is mounted on a deck 1.2 meters above the water surface. This height constraint effectively eliminates reflection points located below or adjacent to the water surface, thereby reducing the interference of surface clutter on the detection results.

[0051] Neighborhood density filtering: Besides height features, surface clutter points are typically sparsely distributed and lack structural continuity. To further remove clutter, a secondary filtering based on the neighborhood density of the point cloud can be performed. For any point in the point cloud... Let the number of its neighborhood points within the radius r be... Then its neighborhood density is defined as:

[0052] (3)

[0053] The radius *r* is 1 meter, determined based on the 3cm resolution of the lidar and the distribution characteristics of the water surface point cloud, to filter out points with a neighborhood density less than a threshold. If the obtained neighborhood point density... If the value is less than 1, the point is considered an isolated clutter point and is removed. Through the combined processing of height threshold filtering and neighborhood density filtering, the interference from water surface reflection points can be significantly reduced, preserving valid target point cloud data.

[0054] Based on the Voxblox hash map framework, the preprocessed aligned point cloud is converted into structured voxel units, constructing a global dynamic voxel map that supports incremental updates. By assigning each voxel three categories of six core attributes—spatial feature attributes, temporal attributes, and state attributes—a foundation is provided for subsequent dynamic voxel discrimination. Further, step S2 includes the following steps:

[0055] S21: Establishing global voxel mesh generation and point mapping relationships:

[0056] Based on the LiDAR hardware parameters and the water surface application scenario, the three-dimensional spatial range of the voxel map is defined. In this invention, the horizontal field of view of the LiDAR is 120°, the maximum detection range is 130m, the X-axis is the horizontal field of view with a range of [−50, 50], the Y-axis is the vertical depth field of view with a range of [0, 130], and the Z-axis is the vertical height field of view with a range of [-1.5, 5], ensuring the complete inclusion of water surface targets.

[0057] A voxel is a cubic grid in space. Its length design needs to balance resolution and computational efficiency. Smaller voxels offer higher precision, but if the number of voxels is too large, it may not meet real-time requirements. Through experimental verification, in this invention, the side length of the voxel is set to 0.2m.

[0058] For each 3D point cloud Assign a unique voxel index This enables a fast mapping relationship between points and voxels. Each voxel records the coordinates of all the point clouds it contains. Subsequent attribute calculations only need to call this mapping table, avoiding repeated traversal of the original point cloud. In actual operation, only non-empty voxels containing point clouds are stored, eliminating the need to traverse all voxels, thus saving computational costs.

[0059] S22: Based on the signed distance function, an incremental update strategy is adopted, requiring only the initial map initialization, and subsequent updates only update the attributes of changing voxels:

[0060] The initial frame provides a baseline for the map and requires initial attribute assignments to all voxels. Each voxel can be represented as a multi-attribute defined state:

[0061] D i To truncate the cumulative value of the TSDF (Signal Distance Function), D i =0 indicates that there is no accumulated information in the voxels;

[0062] W i To truncate the cumulative observation weights of the TSDF (Signal Distance Function), W i =0, meaning the voxel was not observed;

[0063] The most recently occupied timestamp for the current voxel; This indicates that the space was not initially occupied.

[0064] For continuous occupation duration; This indicates that the initial occupancy was not contiguous;

[0065] f is the voxel free state discrimination value, f=0 indicates non-free, f=1 indicates free. Initially, it is non-free by default. If the current voxel satisfies the condition that the voxel's TSDF value is greater than the preset threshold, then f is allowed to determine the free state. , or there may be a point cloud in the current frame. If it falls into that voxel, then it is updated. If a voxel is not free, it is considered a non-free voxel; otherwise, it is considered a free voxel. Set an N value, and for the current non-free voxel, if it was free in the previous N frames, it is determined to be a dynamic voxel. Then, all the voxels determined to be dynamic are collected into a set of dynamic voxels.

[0066] S23: Voxel multidimensional attributes accumulate with frame rate updates:

[0067] First, calculate sdf, which is the difference between the distance from point cloud p to the origin and the distance to its nearest voxel center. This value can be negative. The truncation distance s is set to 0.3m in this invention, which is 10 times the lidar resolution of 3cm, balancing accuracy and anti-interference capability. It is calculated using the following formula:

[0068] (4)

[0069] Cumulative TSDF value and cumulative weight The update formula is as follows:

[0070] (5)

[0071] in This completes the establishment of the voxel TSDF update mechanism. For a currently non-empty voxel, its attributes include... Assigning the current timestamp increases the voxel's occupancy time. The value f, which represents the free state of a voxel, is an important attribute for dynamic target detection. In this invention, the detection of dynamic voxels is defined as follows: if a currently non-empty voxel was in a free state in the previous frame, i.e., f=1, then it is considered a dynamic voxel.

[0072] Furthermore, step S3 includes the following steps:

[0073] S31: Constructing basic point cloud local feature descriptors:

[0074] To ensure the real-time performance of the entire detection framework, this invention combines local and global features based on FPFH. In sparse point cloud scenarios on water surfaces, relying solely on local features is prone to feature fragmentation due to missing point clouds, while relying solely on global features easily leads to the loss of target details. This invention addresses the problem of insufficient feature information in sparse point clouds by combining local and global vectors, enabling the descriptor to maintain stability in more scenarios. The construction of the point cloud descriptor is as follows: Figure 2 As shown. p i The point is the center point of the voxel, n i It is p i The normal vector of a point, p j A point is a point in the neighborhood, n j It is p j The normal vector of a point is used to calculate the feature descriptor. Let's take the normal vector n. i As the axis u, take the AND... (i.e., vector p) i -p j(direction) and normal vector n i The direction perpendicular to both axes u and v is used as another axis v, and finally, an axis w orthogonal to both axes u and v is obtained, thus determining the local reference frame. The parameters of the feature descriptor include angles and vectors. The angle formed with axis u is p i Let the vector formed by the point and the origin of the global coordinate system be n0. The angle formed with n0 is ;n j The angle formed with axis u is α, n j The projection onto the plane formed by the v / w axes makes an angle θ with the w-axis. Local and global vectors together constitute feature elements, enabling the descriptor to remain stable in more scenarios.

[0075] Curvature, like normal vectors, is an important geometric feature characterizing point clouds and exhibits permutation invariance. Point p i The curvature K can be obtained from its neighborhood points. A point p can be found using a KD-tree (K-Dimension Tree). i The M nearest neighbors form an M×3 matrix:

[0076] (6)

[0077] Let be the covariance matrix, and k be the number of neighboring points. Therefore, calculate... The eigenvalues ​​of the minimum eigenvalue Point p i The neighborhood curvature is as follows:

[0078] (7)

[0079] Finally, the point cloud feature descriptor construction is complete:

[0080]

[0081] Where d is the distance between the two points.

[0082] This step uses the voxel core attributes updated in step S3 as input. Based on the voxel attributes and state, and in accordance with the definition, a currently non-empty voxel is considered a dynamic voxel if it was in a free state in the previous frame. A description of dynamic voxel detection is then developed. Step S4 includes the following steps:

[0083] S41: Preliminary dynamic discrimination based on voxel occupancy state and spatiotemporal attributes:

[0084] If the current voxel meets one of the following two conditions, update its most recently occupied timestamp and occupation duration, and treat it as a non-free voxel:

[0085] 1. The TSDF value of the voxel is greater than the preset threshold. In this invention, A value of 0.1 ensures that voxels geometrically close to the target surface are accurately identified as occupied.

[0086] 2. There exists a point cloud in the current frame that falls into this voxel. Conversely, if none of the above conditions are met, then the voxel is in a free state, and f=1.

[0087] Due to the influence of water flow and waves, targets on water are prone to pose drift, causing jitter in the point cloud position of static voxels. Identifying dynamic voxels solely based on TSDF values ​​or timestamps can easily lead to misclassification of such jittery static voxels as dynamic ones. This invention introduces an anti-jitter threshold and spatial consistency constraints, filtering drift interference through a dual 'temporal + spatial' constraint to reduce misclassification. Therefore, an anti-jitter threshold is introduced. , The timestamp of the current frame, if ,in If a voxel has been occupied within two frames, it will not be considered free. If a voxel is not occupied within 5 frames, it is considered to be in a free state. In addition, there is a spatial consistency constraint: before a voxel is identified as free, it is ensured that the cumulative weight W of all voxels in the 3×3×3 neighborhood of the current voxel is greater than 0, indicating that the neighborhood voxels have been observed by the lidar at least once, to prevent unscanned space from being misjudged.

[0088] S42: Using feature descriptors to identify dynamic voxels:

[0089] The covariance descriptor provides information about the correlation between features, is insensitive to pose changes, and is more stable to disturbances. Its calculation method is as follows:

[0090] (8)

[0091] Where N represents the dimension of the original feature vector, and F is the feature vector obtained in S32, while Let C be the average of all eigenvectors within a voxel, and let C be the covariance descriptor of the voxel. The covariance descriptor is symmetric; in voxel feature comparison, the similarity between two matrices is reflected by the distance in space. Riemann space is commonly used to measure symmetric positive definite matrices; it preserves curvature information in local regions and is suitable for point cloud similarity comparison. The Riemann space distance after feature matching can be expressed as:

[0092] (9)

[0093] These are the two covariance matrices that need to be compared. This non-Euclidean space measurement method can better reflect the distance between features and achieve a good balance between evaluation accuracy and computational speed. Subsequently, all non-empty voxels are used as input, and feature extraction is performed to obtain the point cloud feature descriptor of the current frame. The feature descriptor of the previous frame is substituted into formula (9) to obtain the distance between the two feature matrices.

[0094] (10)

[0095] in, Based on a large amount of water and land scene data, the system considers distances greater than a threshold as changes and identifies them as free voxels, while those less than a threshold are considered non-free voxels.

[0096] Furthermore, step S5 includes the following steps:

[0097] S51: Clustering dynamic voxels into complete targets using dynamic voxel clustering and noise filtering algorithms:

[0098] The detected dynamic voxels only represent local regions of the complete moving target, and the complete target contour needs to be obtained through connected component growth. Conventional region growing algorithms only determine neighborhood correlation based on spatial distance, which easily leads to missegmentation of sparse voxels at the target edge and noise voxels. This invention adds two constraints, distance difference and cumulative weight, to the region growing process, and combines them with a voxel number threshold to filter noise clusters, ensuring that the output dynamic voxels correspond to the real moving target. Starting with the dynamic voxel seed selected by S42, a 26-neighbor search (i.e., all neighboring voxels of the voxel in 3D space) is used to merge neighboring voxels that meet the following conditions to form a dynamic voxel cluster: Condition 1: Neighboring voxels are identified as non-empty in the current frame; Condition 2: The TSDF distance difference between the neighboring voxel and the seed voxel is less than 0.2m, ensuring that they belong to the same target surface and avoiding cross-target merging; Condition 3: The cumulative weight W of the neighboring voxels is greater than 2, and they have been observed by the lidar at least twice, excluding isolated voxels caused by single-point cloud noise.

[0099] The dynamic voxel cluster corresponding to the moving target needs to have a certain degree of spatial continuity. Using a region growing algorithm, if the number of voxels in the dynamic voxel cluster is less than the minimum number of voxels in a single cluster... If the number of voxels is greater than or equal to a certain value, it is considered a noise cluster and removed from the dynamic voxel set. These are retained as effective dynamic voxel clusters, with each cluster corresponding to an independent moving target. In this paper... With 5 voxels, each corresponding to a space of 0.04 m³, small clusters of noise can be effectively eliminated.

[0100] S52: Construction of Experimental Platform and Test Dataset:

[0101] All data acquisition and algorithm verification were performed on the same platform, which consisted of a LeiShen 64-line LiDAR with a horizontal viewing angle of 120 degrees, a vertical viewing angle of 11 degrees, and a resolution of 3 centimeters, fully meeting the experimental requirements. The data processing platform was an Intel NUC equipped with an i7-8559U processor and 16GB of memory.

[0102] To make the validation of the algorithm more convincing, a total of 3000 frames of point cloud data were collected on both water and land. In the land scene, the coverage included small objects in complex indoor environments, fast-moving vehicles in road environments, and various uncertain objects in campus environments, thus maximizing the diversity and richness of the experimental dataset. Validation was also performed on publicly available datasets, L-CAS and DOALS, two 3D point cloud datasets primarily composed of pedestrian laser point cloud data from campus scenes. For the water scene, we collected point cloud data of boats of different sizes in a real river environment. To comprehensively evaluate the algorithm's performance in various aspects, we also collected data on boats with different movement speeds, distances, and sizes.

[0103] S53: Experimental evaluation indicators:

[0104] Since our detection method can also segment dynamic point clouds, the Interchange of Units (IoU), commonly used in point cloud segmentation, is employed as a metric. For a single detected object or category, the IoU is defined as:

[0105] (11)

[0106] In this method, TP represents a true positive point, indicating that the point is actually dynamic and has been correctly identified as such; FP represents a false positive point, indicating that the point is actually static but has been mistakenly identified as dynamic; and FN represents a false negative point, indicating that the point is actually dynamic but has been mistakenly identified as static. This method is more accurate than calculating the overlap of volumes, thus making the experimental results more convincing.

[0107] S54: Analysis of Experimental Results in General Scenarios:

[0108] Figure 3 The specific results of moving object segmentation are presented. The visualization results demonstrate the competitiveness of this invention across various application scenarios and diverse detection objects. In highway environments, objects include moving cars, bicycles, and pedestrians, and this method effectively segments these common objects.

[0109] In indoor or densely built-up environments, radar generates high-resolution, dense point clouds, which are more conducive to object detection. In such cases, the present invention achieves satisfactory performance with virtually no omissions or false detections. Figure 5 As shown, the proposed method achieves the best results in both precision and IoU metrics in typical scenarios, outperforming other methods by 10.2% and 6.8% respectively, but is slightly inferior in recall. Regarding runtime, the average runtime per frame of this invention is 77 milliseconds, meeting the requirements for real-time detection.

[0110] S55: Analysis of experimental results in water surface scenarios:

[0111] The above analysis has demonstrated that the proposed method has good detection capabilities in general scenarios. In this step, we will analyze the performance of the algorithm on another water surface dataset. In open water environments, point clouds are relatively sparse, making the detection of moving objects more challenging. During movement, large objects often have numerous overlapping portions in adjacent frames, and it is difficult to detect these overlapping voxels using only simple placeholder information methods, resulting in low IoU. For example... Figure 4 As shown, the designed feature matching algorithm effectively addresses the problem of sparse and difficult-to-detect objects on the water surface, and can detect large moving objects regardless of the angle. The visualization results demonstrate that before approaching the buoy, the vessel is simply detected as moving, but once it begins to drag the buoy, the algorithm also marks the moving buoy as dynamic.

[0112] To better quantify the results and demonstrate the relationship between algorithm performance and distance, objects on the water surface were divided into three different intervals based on distance, and the IoU was calculated for each interval. Experiments were conducted using the metrics in formula (11) on the collected data, and... Figure 6 The results show comparisons with other moving object detection methods in a water surface environment. Compared to conventional scenes, our method leads in every metric in water surface scenes, and the lead increases with distance. An object is considered successfully detected when the detected intersection-union ratio (IUU) is greater than 0.5, and all proposed methods are able to detect this object. As the point cloud distance increases, the time required for map construction also increases, thus increasing the algorithm's runtime on water surfaces; however, a frame rate of 91.2 milliseconds still meets real-time requirements.

[0113] Assembling the above methods and steps into a program and storing it on a hard disk or other non-transitory storage medium constitutes an embodiment of the present invention's "a non-transitory readable recording medium"; while electrically connecting the storage medium to a computer processor and enabling the detection of moving targets on the water surface through data processing constitutes an embodiment of the present invention's "a system for detecting moving targets on the water surface".

[0114] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computers or available storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0115] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0116] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0118] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., 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 detecting a moving target on the water surface, characterized in that, Includes the following steps: S1. The raw point cloud data of the target obtained by the lidar and synchronous inertial measurement unit is filtered and then preprocessed with temporal and spatial alignment. S2. Divide the preprocessed point cloud into voxels, construct a real-time updated global voxel map, and assign different dimensions of observation attributes to each voxel to obtain a preliminary dynamic voxel set from a spatiotemporal perspective. S3. Within the initial set of dynamic voxels, voxel features are obtained using the covariance matrix. Local and global features and surface curvature information are fused to construct a point cloud feature descriptor, which characterizes the geometric feature differences of voxels, distinguishes between dynamic and non-dynamic voxels, and removes non-dynamic voxels from the set of dynamic voxels to obtain the set of corrected dynamic voxels. S4. Using the corrected dynamic voxels as seeds, perform region growth, merge adjacent corrected dynamic voxels, and obtain the moving target; Time alignment is achieved by using a timestamp synchronization module to map all data onto the same time axis, thus realizing the time sequence alignment of multi-sensor data. Spatial alignment is achieved by accurately mapping the LiDAR point cloud of the previous frame to the reference coordinate system of the next frame through pose transformation; filtering includes: height threshold filtering to remove reflection points located below or adjacent to the water surface and neighborhood density filtering to filter reflection points with neighborhood density less than the threshold. Obtaining the initial dynamic voxel set from a spatiotemporal perspective involves the following steps: For each voxel, it can be represented by multiple observed attribute definitions: Each attribute is defined as follows: D i To truncate the cumulative value of the TSDF (Signal Distance Function), D i =0 indicates that there is no accumulated information in the voxels; W i To truncate the cumulative observation weights of the TSDF (Signal Distance Function), W i =0, meaning the voxel was not observed; The most recently occupied timestamp for the current voxel; This indicates that the space was not initially occupied. For continuous occupation duration; This indicates that the initial occupancy was not contiguous; f is the voxel free state discrimination value, f=0 indicates non-free, f=1 indicates free. Initially, it is non-free by default. If the current voxel satisfies the condition that the voxel's TSDF value is greater than the preset threshold, then f is allowed to determine the free state. , or there may be a point cloud in the current frame. If it falls into that voxel, then it is updated. A voxel that has been occupied within two frames is not considered free; a voxel that has not been occupied within five frames is considered free. An N value is set; for a currently non-free voxel, if it was free in the previous N frames, it is considered a dynamic voxel. Additionally, there is a spatial consistency constraint: before being considered a free voxel, it is ensured that the cumulative weight W of all voxels within its 3×3×3 neighborhood is greater than 0, indicating that the neighborhood voxels have been observed by the lidar at least once. Then, all voxels determined to be dynamic are grouped into a set of dynamic voxels. During the region growing process, the number of voxels in the resulting dynamic voxel cluster is less than the minimum number of voxels in a single cluster, achieved through a region growing algorithm. If a cluster is identified as noise, it is removed from the dynamic voxel set.

2. The method for detecting a moving target on the water surface according to claim 1, characterized in that, Point cloud feature descriptors also include point curvature calculation, which supplements the feature dimensions in distorted scenes by utilizing the stable representation characteristics of curvature on the geometric structure of the target.

3. The method for detecting a moving target on the water surface according to claim 2, characterized in that, The differences in the geometric characteristics of voxels, characterized by the point cloud covariance matrix, are reflected by distances in Riemann space.

4. A non-transitory readable recording medium for storing one or more programs containing multiple instructions, characterized in that, When the instruction is executed, the processing circuit will perform the method for detecting a moving target on the water surface as described in any one of claims 1-3.

5. A detection system for moving targets on the water surface, comprising a processing circuit and a memory electrically coupled thereto, characterized in that, The memory is configured to store at least one program, the program containing multiple instructions, and the processing circuit runs the program to execute the method for detecting a moving target on the water surface according to any one of claims 1-3.

Images