Fragment detection method and system fusing physical priori and time sequence convolution

Abstract

The application provides an explosion debris detection method and system fusing physical priori and time sequence convolution. The method acquires an image sequence of an explosion field, a burst core coordinate, a time interval between image frames and camera calibration parameters; candidate debris targets are screened according to morphological features and motion features of the image frames, and candidate debris trajectories are established based on a radial relationship between the burst core and the candidate debris targets; a historical state sequence is constructed according to historical positions and motion states of the candidate debris trajectories, a time sequence convolution network is used to predict debris positions and update the candidate debris trajectories; and effective debris trajectories are screened according to geometric consistency between the candidate debris trajectories and the burst core. The method can improve the accuracy and reliability of debris detection and trajectory tracking.

Description

Technical Field

[0001] This application relates to the field of image processing technology, specifically to a method and system for detecting explosion debris that integrates physical priors and temporal convolution. Background Technology

[0002] In related technologies, the intense flashes, smoke, shock waves, and background clutter generated during the explosion significantly reduce the image signal-to-noise ratio. Debris often appears as small, dark targets composed of a few pixels, making it prone to missed detections and false detections. At the same time, debris is affected by air resistance, gravity, and attitude changes, resulting in nonlinear motion. Traditional tracking methods are prone to accumulated prediction errors, trajectory breaks, and identity switching, and it is difficult to eliminate dynamic interference from the explosion's epicenter. Summary of the Invention

[0003] This application aims to at least partially address one of the technical problems in related technologies. To this end, embodiments of this application propose a method and system for detecting explosive debris that integrates physical priors and temporal convolution, which can improve the accuracy and reliability of debris detection and trajectory tracking.

[0004] According to the first aspect of this application, the explosion debris detection method that integrates physical priors and temporal convolution provided by this application includes the following steps: S1, obtain the image sequence of the explosion field and the coordinates of the blast center; S2, Based on the morphological and motion features of the image frames in the image sequence, candidate fragment targets are selected; S3, Based on the radial relationship between the blast center coordinates and the candidate debris targets, the candidate debris targets in consecutive image frames are associated to establish the candidate debris trajectory; S4. Construct a historical state sequence based on the historical position and motion state of the candidate fragment trajectory, input the historical state sequence into a temporal convolutional network to predict the position of the fragment in subsequent image frames, and associate the candidate fragment target in the subsequent image frames with the predicted position to update the candidate fragment trajectory. S5. Based on the geometric consistency between the candidate fragment trajectories and the coordinates of the blast center, select the effective fragment trajectories from the candidate fragment trajectories.

[0005] In some embodiments, the morphological features include the stability of dark connected regions in the image frame, and the motion features include the changes in pixels in the image frame relative to the background; the step of filtering candidate fragment targets based on the morphological and motion features of image frames in the image sequence includes: Based on the stability of the dark connected regions, morphological candidate regions are extracted from the image frames; Based on the changes of the pixels relative to the background, the moving pixel region is extracted from the image frame to generate a moving foreground mask; Based on the spatial overlap between the morphology candidate region and the moving foreground mask, the morphology candidate region is filtered to obtain the target to be filtered. The target to be filtered is filtered based on its attribute information to obtain the candidate fragment target.

[0006] In some embodiments, a maximum stable extreme value region algorithm is used to extract stable dark connected regions in the image frame to form the morphology candidate region, and a visual background extraction algorithm is used to extract moving pixel regions in the image frame to generate the moving foreground mask. The spatial overlap is determined as the ratio of the number of pixels belonging to the moving foreground mask within the morphology candidate region to the total number of pixels in the morphology candidate region, and the morphology candidate region with a spatial overlap greater than a preset overlap threshold is determined as the target to be filtered. The attribute information includes pixel area and grayscale. Targets whose pixel area is within a preset pixel area range and meets a preset grayscale condition are identified as candidate fragment targets.

[0007] In some embodiments, associating candidate debris targets in consecutive image frames based on the radial relationship between the blast center coordinates and the candidate debris targets to establish candidate debris trajectories includes: The candidate fragment targets in the current image frame are identified as targets to be associated. Based on the relative position of the blast center coordinates and the target to be associated in the image coordinate system, determine the radial direction from the blast center coordinates to the target to be associated. Using the radial direction as a matching reference, candidate associated targets whose displacement direction relative to the target to be associated satisfies the preset direction constraint are selected from the candidate fragment targets in the next image frame; From the candidate associated targets, determine the matching target corresponding to the target to be associated, and establish the association relationship between the target to be associated and the matching target; Based on the correlation in adjacent image frames, candidate fragment targets in consecutive image frames are associated to form candidate fragment trajectories.

[0008] In some embodiments, the step of selecting candidate associated targets from candidate fragment targets in the next image frame whose displacement direction relative to the target to be associated satisfies a preset direction constraint, using the radial direction as a matching reference, includes: Using the position of the target to be associated as a reference and the radial direction as an axis, a fan-shaped physical constraint region is constructed. The search radius of the fan-shaped physical constraint region is determined based on the maximum possible displacement of the fragment, and the angle range of the fan-shaped physical constraint region is determined based on a preset initial tolerance angle. Candidate fragment targets located within the fan-shaped physical constraint region in the next image frame are determined as candidate associated targets. The distance between the candidate associated target and the target to be associated does not exceed the maximum possible displacement, and the angle between the displacement direction from the target to be associated to the candidate associated target and the radial direction does not exceed the preset initial tolerance angle. In some embodiments, the step of associating candidate fragment targets in consecutive image frames to form candidate fragment trajectories based on the association relationship in adjacent image frames includes: Candidate fragment targets in three consecutive image frames are sequentially associated according to the temporal order of the image sequence; When the association between candidate fragment targets in adjacent image frames is established and the candidate fragment targets in three consecutive image frames pass the preset smoothness test, the candidate fragment targets in the three consecutive image frames are associated to form a candidate fragment trajectory, and the candidate fragment trajectory is activated.

[0009] In some embodiments, the step of constructing a historical state sequence based on the historical position and motion state of the initial trajectory of the fragment, inputting the historical state sequence into a temporal convolutional network to predict the position of the fragment in subsequent image frames, and associating the predicted position with candidate fragment targets in subsequent image frames to update the fragment trajectory includes: Extract the centroid coordinates of the candidate fragment targets from L consecutive historical image frames from the candidate fragment trajectory, where L is a preset number of historical frames; The first-order differential velocity is determined based on the difference in centroid coordinates between adjacent historical image frames. The normalized centroid coordinates and the first-order differential velocity corresponding to each historical image frame are stored in a sliding window in chronological order to construct the historical state sequence. The historical state sequence is input into the temporal convolutional network to obtain the predicted centroid coordinates and error covariance of the fragment in the next image frame. Based on the positional deviation between each candidate fragment target in the next image frame and the predicted centroid coordinates and the error covariance, the Mahalanobis distance of each candidate fragment target relative to the predicted centroid coordinates is determined, and the candidate fragment target with the smallest Mahalanobis distance that is less than a preset threshold determined based on the chi-square distribution is determined as the matching target corresponding to the candidate fragment trajectory. The fragment trajectory is updated based on the matching target.

[0010] In some embodiments, the temporal convolutional network includes a causal convolutional structure, a dilated convolutional structure, and a residual connection structure, and adopts a three-layer residual block structure with a convolutional kernel size of 3. Updating the fragment trajectory according to the matching target includes: When the matching target is determined, the actual observation state corresponding to the matching target is added to the sliding window to update the historical state sequence, and the matching target is added to the candidate fragment trajectory; When there are no candidate fragment targets with a Mahalanobis distance less than the preset threshold in the next image frame, the predicted state output by the temporal convolutional network at the previous time step is added to the sliding window as the virtual state at the current time step, and the position of the fragment in subsequent image frames is continued to be predicted. When the corresponding candidate fragment target reappears within five image frames and successfully matches the candidate fragment trajectory, the candidate fragment trajectory is resumed; when no corresponding candidate fragment target is matched after five image frames, the candidate fragment trajectory is terminated.

[0011] In some embodiments, the step of filtering valid debris trajectories from the candidate debris trajectories based on the geometric consistency between the candidate debris trajectories and the coordinates of the blast center, and obtaining the corresponding image coordinate sequence from the valid debris trajectories, includes: From the initial stage of the candidate fragment trajectory, the first M observation points are extracted according to the time order of the image frames, where M is a positive integer not less than 2; The least squares method is used to fit a two-dimensional straight line to the first M observation points to obtain a trajectory straight line that represents the initial motion direction of the candidate fragment trajectory; The orthogonal deviation between the coordinates of the blast center and the trajectory line is defined as the geometric deviation; When the geometric deviation does not exceed the preset geometric tolerance threshold, the candidate fragment trajectory is determined as the effective fragment trajectory; when the geometric deviation exceeds the preset geometric tolerance threshold, the candidate fragment trajectory is determined as background interference trajectory and is eliminated. The real observation points in the effective fragment trajectory are extracted, and the image coordinates of each real observation point are arranged in the time order of the corresponding image frames to obtain the image coordinate sequence.

[0012] In a second aspect of this application, the explosion fragment detection system that integrates physical prior knowledge and temporal convolution provided by this application includes: The data acquisition module is used to acquire image sequences of the explosion field and the coordinates of the blast center; The target detection module is used to filter candidate fragment targets based on the morphological and motion features of image frames in the image sequence; The trajectory initiation module is used to associate candidate debris targets in consecutive image frames based on the radial relationship between the blast center coordinates and the candidate debris targets, so as to establish the initial trajectory of the debris; The trajectory tracking module is used to construct a historical state sequence based on the historical position and motion state of the initial trajectory of the fragment, input the historical state sequence into a temporal convolutional network to predict the position of the fragment in subsequent image frames, and associate the predicted position with candidate fragment targets in subsequent image frames to update the trajectory of the candidate fragment. The trajectory verification module is used to filter out valid fragment trajectories from the candidate fragment trajectories based on the geometric consistency between the candidate fragment trajectories and the coordinates of the blast center.

[0013] In summary, the explosion debris detection method and system integrating physical prior and temporal convolution provided in this application have at least the following advantages: a complete processing flow from image and related parameter acquisition, candidate debris target screening, candidate debris trajectory establishment and updating, effective debris trajectory screening to motion trajectory. Specifically, screening candidate debris targets based on the morphological and motion features of image frames, combined with pixel area and grayscale attributes for filtering, helps reduce false detections caused by strong light, smoke, and static dark spots; associating candidate debris targets in consecutive image frames based on the radial relationship between the blast center coordinates and candidate debris targets helps narrow the candidate association range and reduce erroneous associations; constructing a historical state sequence based on the historical position and motion state of candidate debris trajectories, and using a temporal convolutional network to predict the position of debris in subsequent image frames helps maintain trajectory continuity under nonlinear motion and short-term occlusion conditions; and screening effective debris trajectories based on the geometric consistency between candidate debris trajectories and blast center coordinates helps eliminate background interference trajectories that do not conform to the blast center source dispersion characteristics, thereby improving the accuracy and reliability of debris detection and trajectory tracking. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the explosion fragment detection method that integrates physical priors and temporal convolution in the embodiments of this application.

[0015] Figure 2 This is a schematic diagram illustrating the execution of the explosion fragment detection method that integrates physical priors and temporal convolution in the embodiments of this application.

[0016] Figure 3 This is a schematic diagram illustrating the execution of step S2 in an embodiment of this application.

[0017] Figure 4 This is a schematic diagram illustrating the execution of steps S3 and S4 in the embodiments of this application.

[0018] Figure 5 This is a schematic diagram illustrating the execution of step S5 in an embodiment of this application.

[0019] Figure 6 This is a schematic diagram of the structure of the explosion fragment detection system that integrates physical priors and temporal convolution in the embodiments of this application.

[0020] Figure label: 100. An explosion fragment detection system that integrates physical priors and temporal convolution; 110. Data Acquisition Module; 120. Target Detection Module; 130. Trajectory Initiation Module; 140. Trajectory Tracking Module; 150. Trajectory Verification Module; 160. Trajectory Reconstruction Module. Detailed Implementation

[0021] The embodiments of this application are described in detail below, with examples of these embodiments illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0022] like Figure 1 and Figure 2 As shown, this application provides a method for detecting explosive debris that integrates physical priors and temporal convolution. This method can be applied to static explosion tests or other scenarios requiring high-speed target motion trajectories. The method combines the physical prior of debris scattering outwards from the theoretical blast center, morphological and motion information from consecutive image frames, and the learning ability of temporal convolutional networks for nonlinear motion, sequentially completing data acquisition, debris target detection, trajectory initiation, trajectory tracking, trajectory verification, and measurement.

[0023] The physical priors refer to the physical constraints that can be determined or reasonably expected before target detection and trajectory tracking are performed, based on the formation mechanism and motion laws of explosion fragments. In this embodiment, the physical priors include at least the radial relationship that fragments typically scatter outward from the vicinity of the blast center, and may also include constraints such as the maximum possible displacement of fragments between adjacent image frames and the continuity of fragment motion direction over a short period of time.

[0024] Temporal convolution refers to performing convolution processing on the historical position and motion state of fragments along the image acquisition time direction to extract temporal change features from continuous historical states. Compared with processing only based on the current image frame, temporal convolutional networks can utilize positional changes in multiple historical image frames to predict the position of fragments in subsequent image frames.

[0025] The method includes the following steps: S1, obtain the image sequence of the explosion field and the coordinates of the explosion center.

[0026] S2, Based on the morphological and motion features of the image frames in the image sequence, candidate fragment targets are selected; S3, Based on the radial relationship between the blast center coordinates and the candidate debris targets, the candidate debris targets in consecutive image frames are associated to establish the candidate debris trajectory; S4. Construct a historical state sequence based on the historical position and motion state of the candidate fragment trajectory, input the historical state sequence into a temporal convolutional network to predict the position of the fragment in subsequent image frames, and associate the candidate fragment target in the subsequent image frames with the predicted position to update the candidate fragment trajectory. S5. Based on the geometric consistency between the candidate fragment trajectories and the coordinates of the blast center, select the effective fragment trajectories from the candidate fragment trajectories.

[0027] Specifically, before step S1 begins, the image acquisition module, such as the camera or photographic equipment, has completed installation, focusing, exposure, and frame rate settings before the explosion test begins. In step S1, the image acquisition module continuously acquires images during the explosion, obtaining an image sequence. The total number of image frames in the image sequence is denoted as... , Let be a positive integer; denoted as the th frame in the time sequence. frame, The frame number is 1 to 1. ;No. A frame image can be denoted as The above. , and All of these are data identifiers describing image sequences.

[0028] The blast center coordinates are the predicted location of the blast source center in the image coordinate system, which can be obtained by projecting the measurement results of the field control points onto the image. The blast center coordinates are marked as follows: ,in and These represent the coordinates of the expected explosion source center in the horizontal and vertical directions of the image, respectively. If the initial location obtained is the explosion center in the measurement coordinate system, it is first transformed to the image coordinate system consistent with the detection target using the measurement calibration relationship, so as to ensure that the same coordinate reference is used for subsequent radial judgment.

[0029] The time interval between image frames is determined by the acquisition time of adjacent image frames. The first... The frame acquisition time is recorded as , will the Frame and the The time interval between frames is denoted as This refers to the difference between two acquisition times. When the image acquisition module acquires images at a stable frame rate, all time intervals can be the same; when the image has an independent timestamp or there are unequal acquisition intervals, the actual time interval of each pair of adjacent frames is read separately to avoid replacing the actual acquisition time with a fixed frame rate.

[0030] In step S2, morphological candidate regions and moving pixel regions are extracted for each image frame. Morphological candidate regions are suspected fragment regions obtained based on the target's appearance, used to describe dark connected regions that remain relatively stable under grayscale changes. Dark connected regions refer to sets of pixels with relatively low grayscale compared to the surrounding background and whose adjacent pixels are spatially continuous.

[0031] Moving pixel regions are dynamic pixel sets obtained by comparing the current image with the background model, used to describe image regions that change relative to the background. Only regions that simultaneously satisfy morphological features, motion features, and attributes such as area and grayscale can be output as candidate fragment targets.

[0032] Candidate fragment targets are single-frame target records selected for cross-frame association. Each record includes at least the frame number, the set of pixels in the region, the centroid image coordinates, the pixel area, and the grayscale attribute. The centroid is the geometric center calculated based on the pixel positions within the target region, representing the target's position in the image. Candidate fragment targets are only single-frame selection results and have not yet undergone cross-frame trajectory association and centroid geometric consistency verification.

[0033] In step S3, the coordinates of the blast center are used. As a physical prior, the physical prior refers to the known laws governing explosion motion that can be used to constrain calculations before image processing. This scheme uses the physical prior that debris typically scatters outwards from the blast center. For each candidate debris target in the current frame, the physical prior determined by... The radial direction pointing towards the target. Radial relationship refers to the directional relationship between the theoretical blast center, the current target, and its subsequent displacement direction.

[0034] Subsequently, in the next frame, priority is given to searching for detected targets moving outward along this radial direction, and then candidate debris trajectories are established through the correlation between consecutive image frames. Here, the candidate debris trajectories are observation records of the same target arranged in ascending order of frame number, rather than target regions in a single frame.

[0035] In step S4, the historical centroid positions and the image domain motion states obtained from the differences between adjacent positions are read from the established fragment trajectories and arranged into a historical state sequence in chronological order. The historical state sequence is the network input obtained by arranging the position and motion changes of the same trajectory in the most recent few frames according to the acquisition time.

[0036] Temporal convolutional networks (TCNs) are neural networks that use one-dimensional convolutions along the time dimension to extract motion change patterns from continuous historical states. TCNs output the predicted positions and uncertainties of subsequent frames based on the historical state sequence, and then use these predicted positions to select matching targets from the detection results of subsequent frames.

[0037] When a match is successful, the real observation is added and the trajectory is updated; when there is no match for a short period of time, the trajectory is maintained according to the blind prediction rule to avoid immediate interruption due to smoke obscuring or momentary missed detection.

[0038] In step S5, the geometric consistency of the formed trajectory is verified. For example, the initial direction of motion can be fitted using the observation points in the initial stage of the trajectory, and it can be determined whether the reverse extension of this direction is close to the theoretical blast center.

[0039] Candidate fragment trajectories that meet geometric consistency are considered valid fragment trajectories, while those that do not are discarded as background dynamic interference. The actual observation points within the valid trajectories can be arranged by frame number to form an image coordinate sequence.

[0040] In summary, this application establishes a complete processing flow through steps S1 to S6, encompassing image and related parameter acquisition, candidate debris target screening, candidate debris trajectory establishment and updating, effective debris trajectory screening, and motion trajectory determination. Specifically, screening candidate debris targets based on the morphological and motion features of image frames, combined with pixel area and grayscale attributes for filtering, helps reduce false detections caused by strong light, smoke, and static dark spots. Associating candidate debris targets in consecutive image frames based on the radial relationship between the blast center coordinates and the candidate debris targets helps narrow down the candidate association range and reduce erroneous associations. Constructing a historical state sequence based on the historical position and motion state of candidate debris trajectories, and using a temporal convolutional network to predict the debris's position in subsequent image frames, helps maintain trajectory continuity under nonlinear motion and short-term occlusion conditions. Screening effective debris trajectories based on the geometric consistency between candidate debris trajectories and blast center coordinates helps eliminate background interference trajectories that do not conform to the blast center source dispersion characteristics, giving the output results practical physical quantity significance. Therefore, the accuracy and reliability of debris detection and trajectory tracking can be improved.

[0041] In some embodiments, the morphological features include the stability of dark connected regions in the image frame, and the motion features include the changes in pixels in the image frame relative to the background; the step of filtering candidate fragment targets based on the morphological and motion features of image frames in the image sequence includes: Based on the stability of the dark connected regions, morphological candidate regions are extracted from the image frames; Based on the changes of the pixels relative to the background, the moving pixel region is extracted from the image frame to generate a moving foreground mask; Based on the spatial overlap between the morphology candidate region and the moving foreground mask, the morphology candidate region is filtered to obtain the target to be filtered. The target to be filtered is filtered based on its attribute information to obtain the candidate fragment target.

[0042] Specifically, step S2 can use a morphology detection stream and a motion detection stream to generate candidate information in parallel, and then fuse the two results. Here, "detection stream" refers to a processing branch that takes an image frame as input and generates a class of candidate results independently according to a predetermined processing order, rather than a "flow" in the fluid sense.

[0043] The morphology detection stream uses the grayscale and connected region appearance of a single frame image as its basis, sequentially executing processing branches such as image reading or grayscale conversion, extraction of stable dark connected regions, and output of morphology candidate regions; its output is the morphology candidate region. The motion detection stream uses background changes in continuous images as its basis, sequentially executing processing branches such as background sample establishment or update, comparison of the current pixel with the background, and dynamic foreground determination; its output is the moving pixel region and the moving foreground mask. The two detection streams process images in the same coordinate system and merge during the spatial overlap filtering stage, thus forming a dual-stream candidate detection for candidate fragment targets.

[0044] Morphological features refer to the property that the area, contour, or connectivity of a dark connected region remains relatively stable across multiple grayscale thresholds; connectivity refers to the ability of pixels within a region to connect to each other according to preset adjacency rules. Motion features refer to changes between current image pixels and an established background model sufficient to classify them as foreground elements. The background model is a pixel reference formed based on previous images or background samples, used to represent the image state when no moving target is present. Morphological and motion features describe suspected fragment regions from the perspectives of single-frame morphology and changes relative to the background, respectively, providing a basis for screening candidate fragment targets.

[0045] During the execution of step S2, firstly, the first... frame The data is converted to grayscale or its grayscale channels are read, and dark connected regions are examined at different grayscale thresholds. When a connected region remains relatively stable during threshold changes, it is considered a morphological candidate region. The morphological candidate region is only a suspected fragment region extracted based on morphological features, and has not yet undergone motion feature screening or attribute filtering. Its purpose is to retain small dark regions that are similar in appearance to fragments.

[0046] Secondly, the background model is established or updated based on consecutive images, and the first... Each pixel in the frame is compared with the background information corresponding to that pixel location. Pixels similar to the background are identified as background pixels, while pixels whose differences from the background meet the motion determination criteria are identified as moving pixels. The positions of all moving pixels in the image plane are combined to form a binary motion foreground mask. This motion foreground mask is a binary marker map with the same size as the original image. The foreground value indicates that the corresponding position belongs to a moving pixel, and the background value indicates that the corresponding position belongs to a background pixel. The purpose of the "mask" is to limit subsequent spatial statistics to only focus on the marked motion positions.

[0047] Next, each morphology candidate region is spatially aligned with the moving foreground mask of the same image frame, and the number of pixels within the morphology candidate region covered by the moving foreground mask is counted. If the coverage meets the preset requirements, it means that the region has both a stable dark shape and is moving relative to the background at the current moment, so it is retained as a candidate target; regions with insufficient coverage are usually static dark spots, markers, or other background textures, and can be discarded at this stage.

[0048] Finally, attribute filtering is performed on the targets to be filtered. Attribute information includes at least pixel area and grayscale. Pixel area is the number of pixels contained in the target to be filtered, and grayscale can be the average grayscale value of the target to be filtered. Regions with too small an area are likely to be image noise, while regions with too large an area are likely to be smoke, fire edges, or large occlusions; regions that do not meet the dark color condition are also not considered candidate debris targets. After the above filtering, the centroid image coordinates of each retained region are calculated and written into the detection result of the current frame.

[0049] It should be noted that the shape detection stream and the motion detection stream should use the same image size and coordinate origin. If the image is scaled, cropped, or distorted before detection, the corresponding transformation relationship should be recorded, and the centroid should be uniformly restored to the image coordinate system used for subsequent trajectory association when outputting the centroid to prevent incorrect overlap judgments due to inconsistent coordinate references between the two candidate regions.

[0050] In this embodiment, a stable dark shape is used as the basis for the appearance of the fragments, and relative background changes are used as the basis for dynamics. Spatial overlap and attribute filtering are used to complete the fusion. This process can simultaneously suppress static background areas that only have a dark appearance but do not move, as well as dynamic interference such as firelight and smoke that only have brightness changes but do not have fragment shapes, thereby improving the accuracy and purity of the candidate fragment target set in the input trajectory association stage.

[0051] like Figure 3 As shown, the exploded image sequence is fed into the MSER morphology detection stream and the ViBe motion detection stream, respectively. The former outputs stable dark morphology candidate regions, while the latter outputs a moving foreground mask. The two results are obtained through spatial overlap. To integrate. Figure 3The “physical constraint filtering” in the text is a general representation of the pixel area conditions and grayscale conditions determined by the physical imaging features of the debris. It corresponds to the attribute filtering in this embodiment and is not the radial sector physical constraint used in step S3. Figure 3 The "candidate detection set" shown is the same as the candidate fragment target set of the current image frame. They have the same meaning. Figure 3 by This illustrates digital gating; in the implementation where the spatial overlap is greater than a preset overlap threshold, candidate regions equal to the threshold are not retained; when the processor uses a comparison instruction not less than, the digital threshold can be set to the minimum representable value above the strict threshold, so that the gating result is equal to the threshold value. Consistent. Therefore... Figure 3 This corresponds to the dual-stream candidate detection, spatial overlap screening, and attribute filtering process described in step S2.

[0052] In some embodiments, a maximum stable extreme value region algorithm is used to extract stable dark connected regions in the image frame to form the morphology candidate region, and a visual background extraction algorithm is used to extract moving pixel regions in the image frame to generate the moving foreground mask; the ratio of the number of pixels in the morphology candidate region belonging to the moving foreground mask to the total number of pixels in the morphology candidate region is determined as the spatial overlap, and morphology candidate regions with spatial overlap greater than a preset overlap threshold are determined as the targets to be filtered; the attribute information includes pixel area and grayscale, and targets to be filtered whose pixel area is within a preset pixel area range and meets a preset grayscale condition are determined as the candidate fragment targets.

[0053] Specifically, the morphology detection stream can employ the Maximum Stable Extremum Region (MSER) ​​algorithm. The MSER algorithm performs thresholding on an image across a series of grayscale thresholds, tracking the area change of connected regions as the threshold changes. When a dark connected region shows minimal change within a certain threshold range, it is identified as a maximum stable extremum region and used as a morphology candidate region. This allows for the extraction of relatively stable dark patches without requiring fragments to have a fixed contour template.

[0054] Motion detection streams can employ a visual background extraction algorithm, namely the ViBe algorithm. ViBe maintains a set of background samples for each pixel location in the image, comparing the current pixel value with the background samples at that location; if a sufficient number of samples are similar to the current pixel, it is classified as background; otherwise, it is classified as dynamic foreground. The algorithm can also update the background samples according to a preset update strategy to adapt to slow changes in background brightness. A motion foreground mask of the same size as the original image is generated from the dynamic foreground pixels. .

[0055] For the current number The first frame Candidate regions of each form Statistics show that this area also includes moving foreground masks. The spatial overlap is calculated by dividing the number of pixels in the morphology candidate region by the total number of pixels in that region. The spatial overlap is equal to the ratio of the number of moving foreground pixels in the morphology candidate region to the total number of pixels in the morphology candidate region.

[0056] The spatial overlap can be calculated using the following formula: (1) In equation (1), Indicates the first The spatial overlap of each morphological candidate region is a dimensionless number between 0 and 1. This represents the set of pixels representing the candidate region for this shape. This represents the set of pixels that take the foreground value in the moving foreground mask of the current image frame; the intersection symbol represents pixels that belong to both. Indicates the number of pixels in the intersection; This represents the total number of pixels in the morphology candidate region. The preset overlap threshold is denoted as... These are preset parameters determined before processing begins, selected based on the overlap distribution between real fragment regions and interference regions in labeled representative images. They are maintained during the processing of the same set of images. Unchanged. When Greater than If the condition is met, the corresponding candidate region will be retained as the target to be filtered; otherwise, it will be discarded.

[0057] Area filtering is then performed on the remaining targets to be filtered. The preset minimum pixel area and preset maximum pixel area are denoted as follows: and Both are measured in pixels. The number of pixels in the candidate target is located at... to The area condition is satisfied when the interval is between. Used to eliminate isolated noise. This is used to exclude smoke patches, fire edges, or other large areas that are significantly larger than the debris imaging scale. The specific values ​​of the above parameters can be determined based on camera resolution, target distance, expected debris size, and statistical results of representative images.

[0058] Gray-scale filtering uses either the average gray-scale of the target to be filtered or the gray-scale difference between the candidate target and the local background. When using average gray-scale, the sum of the gray-scale values ​​of all pixels within the candidate region is divided by the number of pixels in that region to obtain the average gray-scale, which is then compared with a preset gray-scale condition. When using local gray-scale difference, the average gray-scale of the candidate region of the target to be filtered is compared with the average gray-scale of its surrounding background ring. Since fragments in this scheme typically appear as small dark patches, targets that meet the preset dark-color condition are identified as candidate fragment targets.

[0059] Each candidate fragment target can have a target record established, which includes the frame number. Target number The set of pixels in the region, centroid coordinates, pixel area, and average gray level. Target number. Different detected targets are identified only within the current frame, and trajectory numbers are assigned after trajectories are formed across subsequent frames. This is to avoid mixing up single-frame target numbers with trajectory numbers.

[0060] In this embodiment, MSER can preserve relatively stable dark connected regions under varying grayscale thresholds within a single frame, while ViBe can provide dynamic foregrounds based on temporal changes. The two establish a clear fusion criterion through spatial overlap, and then use area and grayscale conditions to exclude regions with inconsistent scale or optical properties. This combination helps reduce false detections in explosion images with low signal-to-noise ratios, strong light, and smoke, providing a more reliable detection set for subsequent trajectory initiation.

[0061] In some embodiments, associating candidate debris targets in consecutive image frames based on the radial relationship between the blast center coordinates and the candidate debris targets to establish candidate debris trajectories includes: The candidate fragment targets in the current image frame are identified as targets to be associated. Based on the relative position of the blast center coordinates and the target to be associated in the image coordinate system, determine the radial direction from the blast center coordinates to the target to be associated. Using the radial direction as a matching reference, candidate associated targets whose displacement direction relative to the target to be associated satisfies the preset direction constraint are selected from the candidate fragment targets in the next image frame; From the candidate associated targets, determine the matching target corresponding to the target to be associated, and establish the association relationship between the target to be associated and the matching target; Based on the correlation in adjacent image frames, candidate fragment targets in consecutive image frames are associated to form candidate fragment trajectories.

[0062] Specifically, the target to be associated refers to a candidate fragment target in the current image frame whose corresponding target needs to be found in the next image frame. The candidate fragment targets in the current frame are taken as targets to be associated, and their centroid image coordinates are defined as follows: ,in and These are the pixel coordinates of the centroid in the horizontal and vertical directions of the image, respectively. The theoretical centroid image coordinates are still denoted as... .

[0063] The radial direction from the theoretical blast center to the target to be associated is determined by the following formula: (2) In equation (2), Indicates the direction from the center of the explosion to the first The radial direction vector of the target to be associated in the frame; Indicates the image coordinates of the centroid of the target to be associated; Represents the coordinates of the blast center image.

[0064] In other words, The starting point is located at the theoretical blast center, and the ending point is located at the current target to be correlated. Its direction represents the expected direction in which debris scatters outward from the blast center. When comparing actual directions, it can be used... The unit direction, and The length is not considered as the radial direction itself.

[0065] Read the first For all candidate fragment targets in the frame, denote the centroid of any candidate associated target as . .Depend on point to vector It is the inter-frame displacement of the candidate target relative to the target to be associated in the image domain. The unit of this quantity is pixels. If its motion state is expressed in one-frame intervals, it can be called pixels / frame. It is only used for image domain association and is not equivalent to the actual velocity obtained by measurement.

[0066] Will Direction and radial vector The directions of the two targets are compared. When the included angle satisfies the preset directional constraint and the candidate target is located radially outward of the target to be associated, the target is added to the candidate associated target set. Targets that do not meet the directional constraint do not participate in this matching, thereby excluding interfering targets that are obviously moving towards the center of the explosion or have excessive lateral deviation.

[0067] When determining a matching target from the candidate set of associated targets, it can first be sorted by directional deviation from smallest to largest. When the directional deviations are close, the centroid distance, area change, or grayscale change can be compared, and the target with the smallest overall difference can be selected. The area and grayscale mentioned above are only used as auxiliary criteria within the same candidate set and do not change the processing logic that uses radial relationship as the main matching benchmark. If there is no candidate target that satisfies the directional constraint, the target to be associated will not be associated in this frame.

[0068] When multiple targets to be associated simultaneously point to the same detected target in the next frame, their respective directional deviations and centroid distances are compared. Only the pair with the lowest cost retains the association relationship, while other targets to be associated continue to select other candidate targets or remain unmatched. This ensures that a single-frame detected target belongs to at most one trajectory at any given time.

[0069] This embodiment transforms the physical law of debris scattering outward from the theoretical blast center into a directional constraint in the image coordinate system, changing the cross-frame search from a directionless neighborhood search to a radially constrained search. When there are many debris, small target sizes, and other moving areas in the background, this constraint can narrow down the candidate range and reduce the possibility of erroneous associations between adjacent targets.

[0070] In some embodiments, the step of selecting candidate associated targets from candidate fragment targets in the next image frame whose displacement direction relative to the target to be associated satisfies a preset direction constraint, using the radial direction as a matching reference, includes: Using the position of the target to be associated as a reference and the radial direction as the axis, a fan-shaped physical constraint region is constructed. The search radius of the fan-shaped physical constraint region is determined according to the maximum possible displacement of the fragment, and the angle range of the fan-shaped physical constraint region is determined according to the preset initial tolerance angle. The candidate fragment target located within the fan-shaped physical constraint region in the next image frame is determined as the candidate associated target, wherein the distance between the candidate associated target and the target to be associated does not exceed the maximum possible displacement, and the angle between the displacement direction from the target to be associated to the candidate associated target and the radial direction does not exceed the preset initial tolerance angle.

[0071] In some embodiments, the step of associating candidate fragment targets in consecutive image frames to form candidate fragment trajectories based on the association relationship in adjacent image frames includes: Candidate fragment targets in three consecutive image frames are sequentially associated according to the temporal order of the image sequence; When the association between candidate fragment targets in adjacent image frames is established and the candidate fragment targets in three consecutive image frames pass the preset smoothness test, the candidate fragment targets in the three consecutive image frames are associated to form an initial fragment trajectory, and the initial fragment trajectory is activated.

[0072] Specifically, a sector-shaped physical constraint region is constructed in the image plane of the (t+1)th frame, with the centroid pt of the target to be associated as the sector vertex (i.e., the reference) and the direction of the radial vector dt as the sector axis. This sector-shaped region is jointly defined by the search radius and the angle range, and it is not required to traverse all detected targets in the entire image.

[0073] The search radius is calculated using the maximum possible displacement of the fragment. . This is a preset parameter that represents the maximum image displacement that a fragment may produce between a pair of adjacent image frames, in pixels.

[0074] in, The calculation can be performed based on the maximum possible physical velocity of the fragments, the time interval between adjacent frames, and the calibration ratio, while allowing for margins for coverage measurement errors and motion fluctuations; alternatively, it can be determined based on the statistical results of inter-frame displacements with labeled real trajectories. In this embodiment, This indicates the maximum inter-frame image displacement, and should not be confused with actual displacement in physical space.

[0075] The sector angle range adopts a preset initial tolerance angle. . These are the angle parameters determined before processing begins, allowing for a certain deviation between the actual debris direction and the ideal radial direction. This deviation can originate from errors in the blast center positioning, image measurement errors, air resistance, or the initial attitude of the debris. The smaller the value, the more stringent the direction selection; The larger the value, the wider the candidate coverage.

[0076] When the j-th detected target in frame t+1 becomes a candidate associated target, the following distance and orientation conditions should be met simultaneously: (3) In equation (3), Indicates by the first The image domain inter-frame displacement from the target to be associated in the frame to the candidate target to be associated in the next frame; This indicates the pixel length of the displacement; Indicates the maximum possible displacement; This represents the radial vector pointing from the theoretical blast center to the target to be associated. This represents the angle between the inter-frame displacement direction and the radial direction; This indicates the preset initial tolerance angle.

[0077] In other words, Formula 3 is used to indicate that the candidate target should neither be too far from the current target nor deviate too much from the expected direction from the center of the explosion outwards.

[0078] like Figure 4 As shown on the left, the "center coordinates" in the figure represent the coordinates of the blast center. The calculation benchmark for the allowable positioning error area is still the blast center coordinates described in this article. The static radial constraint sector region in the figure corresponds to... With axis, The search radius and with A sector-shaped physical constraint region that limits the initial angle range. Figure 4 The actual displacement vector of the annotation Indicates the first The image domain inter-frame displacement of each candidate associated target relative to the target to be associated, its technical meaning is the same as that in equation (3). Same, unit is pixels; Unlike the physical velocity vector obtained after measurement and calibration in equation (8) By making the above distinction, we can avoid mixing up candidate displacements in the image domain with physical velocities in the measurement domain.

[0079] The trajectory is initially examined in chronological order across at least three consecutive image frames. For example, an association is established between the first and second frames, and then between the second and third frames. Only when the three detected targets are temporally consecutive and the two associations point to the same target chain will the smoothness test proceed.

[0080] The smoothness test is used to determine whether two consecutive inter-frame displacements in the image domain conform to the characteristics of continuous motion over a short period of time. It compares the changes in direction and length between the two displacements: if the abrupt change in direction exceeds a preset threshold, or the change in displacement length exceeds a preset threshold, the combination of these three points is considered unsmooth; otherwise, the test is considered passed. Both thresholds are preset parameters, and their values ​​can be determined from the statistical results of short-term changes in real fragment trajectories.

[0081] When targets in three consecutive image frames are associated and pass the smoothness test, the three targets are arranged into an initial fragment trajectory according to the frame number. ,in The trajectory is assigned a number, and its status is changed from "candidate" to "activated". Activated trajectories enter the temporal convolutional prediction and tracking process; target combinations that fail the test are not activated and can be retried in subsequent frames.

[0082] This embodiment simultaneously restricts the search distance and direction of movement by using a fan-shaped physical constraint region. This allows for narrowing the matching range by utilizing the explosion scattering pattern when the trajectory is just beginning and there is not yet sufficient historical data. The correlation and smoothness check of three consecutive frames further increases the requirement for temporal continuity, which can reduce the possibility of single-frame noise, instantaneous flashes, or occasional false detections being incorrectly activated as fragmented trajectories, thereby improving the initial reliability of subsequent prediction and tracking.

[0083] In some embodiments, the step of constructing a historical state sequence based on the historical position and motion state of the candidate debris trajectory, inputting the historical state sequence into a temporal convolutional network to predict the position of the debris in subsequent image frames, and associating the predicted position with candidate debris targets in subsequent image frames to update the candidate debris trajectory includes: Extract the centroid coordinates of the candidate fragment targets from L consecutive historical image frames from the candidate fragment trajectory, where L is a preset number of historical frames; The first-order differential velocity is determined based on the difference in centroid coordinates between adjacent historical image frames. The normalized centroid coordinates and the first-order differential velocity corresponding to each historical image frame are stored in a sliding window in chronological order to construct the historical state sequence. The historical state sequence is input into the temporal convolutional network to obtain the predicted centroid coordinates and error covariance of the fragment in the next image frame. Based on the positional deviation between each candidate fragment target in the next image frame and the predicted centroid coordinates and the error covariance, the Mahalanobis distance of each candidate fragment target relative to the predicted centroid coordinates is determined, and the candidate fragment target with the smallest Mahalanobis distance that is less than a preset threshold determined based on the chi-square distribution is determined as the matching target corresponding to the fragment trajectory. The fragment trajectory is updated based on the matching target.

[0084] Specifically, for trajectory numbered trajectory From its most recent continuous Extract the centroid of the detected target from historical image frames. The preset number of historical frames is the number of states that the sliding window can accommodate; The value of should be consistent with the input length used during the training of the temporal convolutional network and remain unchanged during a single model inference process.

[0085] Wherein, for historical image frames that match candidate debris targets, the state position is the centroid coordinates of the corresponding candidate debris target; for historical image frames that do not match candidate debris targets, the state position is the predicted centroid coordinates output by the temporal convolutional network for the corresponding image frame. Normalize the coordinates of each historical centroid. For example, the horizontal coordinates can be normalized using the image width, and the vertical coordinates can be normalized using the image height, so that the position values ​​in images of different resolutions are within a similar numerical range. The image width and image height refer to the number of pixel columns and pixel rows in the image entering the detection and tracking process, respectively; if the image has been cropped or scaled, the actual size of the processed image is used to maintain consistency between training and inference.

[0086] The difference in centroid coordinates between adjacent historical frames is used as the first-order differential velocity. This first-order difference describes the coordinate change of the target in the image plane after each frame. Before time and calibration scaling, its unit is pixels / frame. After normalization, it is combined with the normalized centroid coordinates to obtain the state corresponding to each historical frame. The state contains information on both "where the target is located" and "where the target recently moved to and how much it moved".

[0087] Recently Each state is stored in a sliding window from earliest to latest according to the collection time, resulting in a historical state sequence. Let the time index corresponding to the last state in this historical state sequence be denoted as . Then the network prediction object is the first Each time position is determined. For each new true observation state, the oldest state in the window is deleted and the latest state is added to the end, thus maintaining the window length at [value missing]. . It is a dynamic state that updates with the trajectory, not a preset parameter.

[0088] Temporal convolutional network reading Then, the predicted centroid coordinates of the next image frame are obtained through the position output branch. The error covariance corresponding to the predicted position is obtained through the uncertainty output branch. The error covariance is a two-dimensional matrix used to describe the prediction error scales in the horizontal and vertical directions, as well as the correlation between the two directions; the more uncertain the prediction, the larger the permissible correlation region can be.

[0089] For the first in the next image frame There are 1 candidate fragment target, whose centroid coordinates are: Compare the centroid with the predicted centroid. Deviation and error covariance between Substituting the squared Mahalanobis distance, the relationship is calculated as follows: (4) In equation (4), Indicates the first The squared Mahalanobis distance of each detected target relative to its predicted location; Indicates the next image frame number The two-dimensional centroid coordinates of the detected target; This represents the two-dimensional centroid coordinates predicted by the temporal convolutional network; the difference between the two represents the positional bias; the transpose superscript in the formula... Indicates vector transpose; Represents the prediction error covariance matrix; This represents the inverse of the matrix.

[0090] In other words, the squared Mahalanobis distance is a measure of how close the detected target is to the predicted position under the prediction error scale; the smaller the value, the higher the degree of matching.

[0091] In this embodiment, This is used to represent the squared Mahalanobis distance. Since the centroid position includes both horizontal and vertical dimensions, a chi-square distribution threshold with 2 degrees of freedom is used for the gating threshold. The preset confidence level is denoted as... , Preset parameters between 0 and 1; based on The obtained two-dimensional chi-square threshold is denoted as Keep it for now. Targets smaller than this threshold are then selected from them. The smallest as the trajectory The matching target in the next image frame. For example, When the value is 0.95, the two-dimensional chi-square threshold is approximately 5.991.

[0092] If multiple trajectories simultaneously select the same detection target, their squared Mahalanobis distances are compared. Only the trajectory with the smallest distance is assigned the target, while the other trajectories are treated as unmatched or their suboptimal candidates are further examined. This maintains the one-to-one correspondence between trajectories and detection targets.

[0093] In this embodiment, the historical state sequence composed of normalized position and first-order difference can provide continuous motion information to the temporal convolutional network; the joint output of predicted position and error covariance allows the correlation threshold to vary with prediction uncertainty; squared Mahalanobis distance gating further converts position deviation into comparable statistics. This improves the cross-frame matching stability under nonlinear motion, scale differences, and local occlusion conditions, and reduces the probability of trajectory breakage or crosstalk.

[0094] Figure 4 Strict tolerance angle in "Adaptive shrinkage" represents an optional search range adjustment method during the tracking phase. When the initial debris trajectory is activated, and the predicted direction of the TCN's continuous output is stable and the prediction error covariance decreases, the search range can be adjusted within the specified range. Within the range, the candidate search angle is changed from the initial tolerance angle. Shrink to strict tolerance angle When the prediction error increases, the target enters a blind prediction state, or tracking resumes, the search angle remains or recovers. This optional adjustment does not change the... The basic association conditions for establishing the initial sector-shaped physical constraint region have the technical effect of further narrowing the candidate range by utilizing the existing trajectory direction and prediction uncertainty, thereby reducing the probability of erroneous association between adjacent fragments. Figure 4 The "sliding prediction position" on the right corresponds to the predicted centroid coordinates output by the TCN. "Uncertainty estimation" corresponds to error covariance "Adaptive Mahalanobis distance gating" means utilizing and two-dimensional chi-square threshold pair Implement gating. Figure 4 The target disappearance area and predicted trajectory at the bottom are used to represent the blind prediction, reassociation, and trajectory recovery process for up to five image frames.

[0095] It should be noted that the first-order difference velocity in the image domain used in step S4 is the coordinate difference between adjacent state positions in the image coordinate system, used for trajectory prediction in the temporal convolutional network, and its unit can be expressed as pixels / frame. The first-order difference velocity in the image domain is used for trajectory prediction and measurement result output.

[0096] In some embodiments, the temporal convolutional network includes a causal convolutional structure, a dilated convolutional structure, and a residual connection structure, and adopts a three-layer residual block structure with a convolutional kernel size of 3.

[0097] Updating the fragment trajectory according to the matching target includes: When the matching target is determined, the actual observation state corresponding to the matching target is added to the sliding window to update the historical state sequence, and the matching target is added to the candidate fragment trajectory; When there are no candidate fragment targets with a Mahalanobis distance less than the preset threshold in the next image frame, the predicted state output by the temporal convolutional network at the previous time step is added to the sliding window as the virtual state at the current time step, and the position of the fragment in subsequent image frames is continued to be predicted. When the corresponding candidate fragment target reappears within five image frames and successfully matches the candidate fragment trajectory, the candidate fragment trajectory is resumed; when no corresponding candidate fragment target is matched after five image frames, the candidate fragment trajectory is terminated.

[0098] Specifically, the temporal convolutional network consists of three residual blocks, each of which includes causal convolution, dilated convolution, nonlinear activation, and residual connection, with a kernel size of 3. The "three residual blocks" and "kernel size of 3" are network structure parameters determined in this scheme. After training, the network is loaded with fixed weights for trajectory prediction.

[0099] Causal convolution ensures that the output at time t only uses the historical state at time t and before, without reading information from future image frames, making it suitable for frame-by-frame processing according to the image acquisition order. Dilated convolution expands the temporal receptive field by setting an interval between convolution sampling positions, allowing networks with a limited number of layers to utilize both recent and distant historical states simultaneously. Residual connections add the residual block input to the convolution transformation result or add it after dimension matching to mitigate gradient decay during deep network training and preserve original motion information.

[0100] In one exemplary configuration, the inflation factors of the three residual blocks can be set to 1, 2, and 4, respectively. The network ends with a position output branch and an uncertainty output branch. The position output branch outputs the two coordinate components of the centroid of the next frame; the uncertainty output branch outputs the parameters required to construct the two-dimensional error covariance. To enable the error covariance to be used for squared Mahalanobis distance calculation, positive constraints can be imposed on the diagonal elements, or a symmetric positive definite matrix can be constructed from the lower triangular factors output by the network. This construction process can be implemented using matrix operations in existing deep learning frameworks.

[0101] When the next image frame exists, it passes through the chi-square gate and When the smallest matching target is found, the true centroid of that target is read, and the corresponding first-order difference is calculated to form the true observation state. The earliest state in the sliding window is deleted, and this true observation state is added. The matching target and its frame number are then added to the trajectory. At this point, the trajectory status remains normal tracking, and the blind prediction count is reset to zero.

[0102] When there is no detected target with a squared Mahalanobis distance less than a preset threshold in frame t+1, the trajectory is not terminated immediately. Instead, the predicted position output from the network at the previous moment and the motion state formed by the predicted position are added to the sliding window as the virtual state at the current moment. The virtual state should be marked with virtual markers in the trajectory record to distinguish it from the real observation point; its purpose is to retain continuous time input for the next prediction, rather than directly using the virtual point as the final measurement result.

[0103] Set the blind prediction count after entering blind prediction. , This is a dynamic state that updates with the number of consecutive unmatched frames. If no matching target is found after one frame, Add 1. The maximum blind prediction lifetime is denoted as In this embodiment When the target reappears within a range of no more than five image frames, and is again selected using squared Mahalanobis distance gating and minimum distance selection, the reappearing real observation is added to the trajectory, and the range is reset. Then resume normal tracking. During recovery, the virtual state of the same frame can be replaced with real observations to keep the historical window consistent with the real detection results.

[0104] When the number of consecutive unmatched frames reaches the maximum blind prediction lifetime When the trajectory is terminated, no further predictions will be made. The terminated trajectory can still retain previous real observation records for geometric verification.

[0105] During network training, a length of [length] can be constructed from the confirmed fragment trajectories. The training data uses historical state samples and the ground truth centroid of the next frame as positional supervision information. The coordinate normalization method and the number of historical frames in the training data are also considered. The order of state components should be consistent with the actual inference. The specific number of training rounds, learning rate, and batch size are part of the model training configuration and can be determined based on the sample size and convergence. They are not considered as mandatory limitations of this implementation method.

[0106] In this embodiment, causal convolution ensures that frame-by-frame prediction does not use future information, dilated convolution expands the historical perception range under a limited number of layers, and residual connections are beneficial for stable network training. Location branches and uncertainty branches provide prediction centers and error scales for subsequent statistical associations. The real state update, combined with blind prediction of up to five frames, can maintain trajectory continuity in the event of short-term occlusion or missed detection, while the lifetime cap prevents the infinite extension of erroneous trajectories caused by long-term use of virtual states.

[0107] In some embodiments, the step of filtering valid debris trajectories from the candidate debris trajectories based on the geometric consistency between the candidate debris trajectories and the coordinates of the blast center, and obtaining the corresponding image coordinate sequence from the valid debris trajectories, includes: From the initial stage of the candidate fragment trajectory, the first M observation points are extracted according to the time order of the image frames, where M is the preset number of observation points and M is a positive integer not less than 2; The least squares method is used to fit a two-dimensional straight line to the first M observation points to obtain a trajectory straight line that represents the initial motion direction of the candidate fragment trajectory; The orthogonal deviation between the coordinates of the blast center and the trajectory line is defined as the geometric deviation; When the geometric deviation does not exceed the preset geometric tolerance threshold, the candidate fragment trajectory is determined as the valid fragment trajectory; when the geometric deviation exceeds the preset geometric tolerance threshold, the candidate fragment trajectory is determined as a background interference trajectory and is discarded.

[0108] Specifically, for the trajectory to be verified Read the initial stage in ascending order of frame number. There are [number] real observation points. Among them, The number of the trajectory to be verified; The preset number of observation points should not exceed the number of existing real observation points for the trajectory, and should be able to support two-dimensional straight line fitting; virtual predicted points are not included in the initial count. One observation point.

[0109] The initial stage of the trajectory was chosen for verification because the direction of motion of debris immediately after leaving the explosion source usually best reflects the characteristics of the source, which disperses outward from the epicenter. Subsequent stages may be affected by air resistance, gravity, attitude changes, or collisions, resulting in more pronounced curvature. Therefore, it is not appropriate to simply use all points along the entire long trajectory to represent the initial direction.

[0110] Using the least squares method for the previous A two-dimensional straight line is fitted using the image coordinates of each observation point. To accommodate both near-vertical and near-horizontal trajectories, the fitted straight line is expressed using a general formula: (5) In equation (5), Representing the trajectory The initial trajectory is a straight line; and Represents image coordinate variables; , and Indicates from the previous The parameters of the straight line obtained by least squares fitting of observation points, and and Not both are 0. Least squares fitting means determining a straight line such that the first... The overall deviation of each observation point from the line should be minimized, so that... Characterizes the initial direction of motion of the trajectory.

[0111] The coordinates of the explosion center Substituting the orthogonal distance relationship from the point to the line, we obtain the geometric deviation: (6) In equation (6), Indicates the trajectory from the center of the explosion. Initial fitted line The orthogonal geometric deviation, in pixels; and The image coordinates represent the theoretical blast center; , and is the line parameter in equation (5); the absolute value ensures that the distance is non-negative; the denominator is used to eliminate the influence of the overall scaling of the line parameter on the distance.

[0112] Geometric deviation is the shortest image distance obtained by drawing a perpendicular line from the center of the explosion to the initial trajectory direction.

[0113] The preset geometric tolerance threshold is denoted as The unit is pixels, and it is a parameter determined before processing. Its value can be determined by comprehensively considering the theoretical center-of-explosion positioning error, centroid detection error, initial trajectory fitting error, and the possible initial deflection of debris. When No more than At that time, the trajectory was considered The reverse extension can converge near the epicenter, identifying it as a valid debris trajectory; when Exceed At that time, it was considered that it did not conform to the single-source dispersion relationship, so it was marked as a background interference trajectory and removed.

[0114] In this embodiment, only the actual observation points are extracted from the verified valid trajectories, forming an image coordinate sequence according to the temporal order of the corresponding image frames. The trajectory status, verification results, and geometric deviation values ​​can be saved together with the image coordinate sequence for easy subsequent inspection and screening of causes. Furthermore, by utilizing the geometric feature that the initial directions of the actual debris trajectories converge in the opposite direction to the vicinity of the theoretical blast center, a source consistency check is performed on the trajectories that already have temporal continuity. This can eliminate dynamic interference such as soil splashes, site debris, and equipment vibration edges, which, although forming continuous motion trajectories, do not originate from the theoretical blast center, thus improving the authenticity of the trajectories entering the measurement stage.

[0115] like Figure 5 As shown, the results of the single-source convergence verification are illustrated using two trajectories to be verified. For the first... Trajectory, using the previous The fitted trajectory straight line at each real observation point And calculate the theoretical core. arrive orthogonal geometric deviation . Figure 5 In It is to satisfy A specific numerical case of the retention condition does not exclude... The data is retained according to the "not exceeding" threshold; (as shown in the figure) This indicates that the geometric deviation exceeds a threshold, and the corresponding trajectory is identified as background interference and removed. Therefore, Figure 5 The two examples shown correspond to valid fragment trajectories and invalid background interference trajectories, respectively, which correspond to the geometric consistency screening rules in step S5.

[0116] When implementing this solution, the following complete process can be followed.

[0117] First, before the explosion test, image acquisition equipment and calibration control points are set up to determine the measurement coordinate system, the location of the blast center, and the imaging area of ​​the image acquisition equipment. Calibration images are acquired, and camera calibration parameters are calculated based on the image coordinates and measurement coordinates of the control points. The camera's focal length, resolution, position, and shooting direction are checked to ensure that the calibration conditions are consistent with the actual test conditions.

[0118] Secondly, several background image frames were acquired before the explosion to establish the ViBe background sample model. The image acquisition device was activated to continuously record the explosion and debris scattering process at a preset frame rate, and the acquisition timestamps corresponding to each image frame were saved.

[0119] Subsequently, the image sequence was sorted by time and its validity was checked, and the image resolution, total number of image frames, acquisition time of each frame, and coordinates of the blast center were read.

[0120] For each image frame, the MSER algorithm is used to extract stable dark connected regions to form a set of morphological candidate regions; the ViBe algorithm is used to compare the current pixel with the background sample to generate a moving foreground mask; the spatial overlap between each morphological candidate region and the moving foreground mask is calculated, and candidate targets with the required overlap are retained; then attribute filtering is performed based on pixel area and grayscale conditions to obtain a set of candidate fragment targets for the current image frame.

[0121] For newly emerging candidate debris targets, the radial direction is determined based on the blast center coordinates and the current target's centroid coordinates. A fan-shaped physical constraint region is constructed with the current target as the vertex, and candidate associated targets that meet the distance and direction conditions are selected in the next image frame. Targets in three consecutive image frames are associated sequentially, and direction and displacement smoothness checks are performed. When the association is successful in all three frames, the initial debris trajectory is formed and activated.

[0122] For activated trajectories, extract the most recent one. The centroid coordinates and first-order difference velocities in the image domain from each historical image frame are normalized and stored in a sliding window to construct a historical state sequence. The historical state sequence is then input into a TCN containing causal convolution, dilated convolution, and residual connections. The predicted centroid coordinates for the next frame are obtained from the position output branch, and the prediction error covariance is obtained from the uncertainty output branch.

[0123] In the next image frame, the squared Mahalanobis distance of each detected target relative to the predicted centroid coordinates is calculated and compared with a chi-square distribution threshold. The detected target that meets the threshold and has the smallest squared Mahalanobis distance is identified as the matching target. When a match is successful, the sliding window and debris trajectory are updated using the actual observation state, and subsequent positions are predicted.

[0124] When no target is detected that meets the threshold, blind prediction is triggered. The predicted state output by the TCN is added to the sliding window as a virtual state, and prediction of subsequent positions continues. If the fragment reappears within five image frames and falls into the predicted area, the trajectory is recovered using the real observation state; if no match is found within five frames, the trajectory is terminated.

[0125] After completing trajectory tracking, extract the initial stage of each candidate trajectory. For each real observation point, a trajectory line is fitted using the least squares method, and the orthogonal deviation from the blast center to this trajectory line is calculated. Trajectories with orthogonal deviations not exceeding the geometric tolerance threshold are retained as valid debris trajectories, while trajectories exceeding the threshold are discarded as non-core source background interference trajectories, thus obtaining the valid debris trajectories.

[0126] The above process enables a complete detection process, from high-speed photographic image acquisition of the explosion site, debris detection, trajectory initiation, continuous TCN tracking, short-term occlusion recovery, and geometric verification.

[0127] It needs to be made clear that, Figure 1 The flowchart illustrating the explosion fragment detection method integrating physical priors and temporal convolution shown in the document is merely a representation of the logical order provided by one of the many embodiments of this application. This order is intended to clearly illustrate the execution flow of the method in a specific application scenario. However, in practical applications, the flexibility and scalability of this application allow those skilled in the art to adjust the order of the above steps according to different actual needs and environmental conditions, such as having parallel or concurrent relationships between some steps, which will not be elaborated upon here.

[0128] In addition, it should be noted that, Figures 2 to 5 This is only used to illustrate the execution process of each processing step in the embodiments of this application. Wherein, Figure 2 The overall execution process of the explosion fragment detection method is shown. Figure 3 The execution process of step S2 is shown. Figure 4 The execution process of steps S3 and S4 is shown. Figure 5 The execution process of step S5 is illustrated. The processing order, number of targets, trajectory shape, area range, and parameter relationships shown in the figure are merely illustrative for the purpose of understanding the technical solution of this application and do not constitute a limitation on the specific implementation of each processing step or the scope of protection of this application.

[0129] Based on the same inventive concept, such as Figure 6As shown, this application also provides an explosion fragment detection system 100 that integrates physical priors and temporal convolution, which includes a data acquisition module 110, a target detection module 120, a trajectory initiation module 130, a trajectory tracking module 140, a trajectory verification module 150, and a trajectory reconstruction module 160. The data acquisition module 110, target detection module 120, trajectory initiation module 130, trajectory tracking module 140, and trajectory verification module 150 are electrically connected via a communication bus.

[0130] The data acquisition module 110 is used to acquire the image sequence of the explosion field, the coordinates of the explosion center, the time interval between image frames, and camera calibration parameters; the target detection module 120 is used to filter candidate debris targets based on the morphological and motion features of the image frames in the image sequence; the trajectory initiation module 130 is used to associate candidate debris targets in consecutive image frames based on the radial relationship between the explosion center coordinates and the candidate debris targets to establish candidate debris trajectories; the trajectory tracking module 140 is used to construct a historical state sequence based on the historical position and motion state of the candidate debris trajectories, input the historical state sequence into a temporal convolutional network to predict the position of the debris in subsequent image frames, and associate candidate debris targets in subsequent image frames based on the predicted position to update the candidate debris trajectories; the trajectory verification module 150 is used to filter valid debris trajectories from the candidate debris trajectories based on the geometric consistency between the candidate debris trajectories and the explosion center coordinates.

[0131] Specifically, the system consists of an image acquisition device, a processor, a memory, and a data interface. The image acquisition device is used to acquire images of the blast site; the data interface is used to receive image sequences, timestamps, blast center coordinates, and calibration parameters; the memory is used to store program instructions, model parameters, intermediate detection results, trajectory records, and measurement results; when the processor executes program instructions, it implements the data acquisition module 110, target detection module 120, trajectory initiation module 130, trajectory tracking module 140, and trajectory verification module 150. Each module can be implemented by a software functional unit, or by a processor, image processor, or other hardware and software working together.

[0132] The data acquisition module 110 receives image sequences and establishes a frame number and acquisition time for each image frame; it receives the coordinates of the blast center and confirms its coordinate system; it receives camera calibration parameters, control point parameters, or coordinate mapping matrices. This module also checks the integrity of image size, frame order, and timestamps, and sends a unified data record to the target detection module 120. Its output includes at least the image frame, frame number, acquisition time, theoretical blast center image coordinates, and camera calibration parameters.

[0133] The target detection module 120 performs MSER morphological candidate extraction and ViBe motion foreground extraction on each image frame, calculates spatial overlap, and filters based on pixel area and grayscale attributes. This module outputs the target number, region pixels, centroid coordinates, area, and grayscale for each candidate fragment target within the current frame. The target detection module 120 is not responsible for confirming the same fragment across frames; its output serves as the observation set for the trajectory initiation module 130 and the trajectory tracking module 140.

[0134] The trajectory initiation module 130 reads the theoretical blast center image coordinates and the detection set of adjacent image frames, constructs a radial fan-shaped physical constraint region for the target to be associated in the current frame, and filters candidate associated targets that meet the requirements in both distance and direction in the next frame. This module checks the association relationship and smoothness of three consecutive image frames in chronological order, assigns trajectory numbers after passing the check, and outputs the activated initial fragment trajectory.

[0135] The trajectory tracking module 140 maintains a sliding window of length L, a real or virtual state marker, a predicted position, an error covariance, and a blind prediction count for each active trajectory. This module calls a temporal convolutional network to generate the prediction result for the next frame, calculates the squared Mahalanobis distance for each detected target, performs two-dimensional chi-square gating, and minimum distance selection. When a match is successful, the trajectory is updated using real observations; when no match is found, a maximum of five frames of blind prediction are performed, and the trajectory state is updated according to recovery or termination conditions.

[0136] The trajectory verification module 150 reads the coordinates of the first M real observation points and the theoretical blast center image of the trajectory to be verified, fits the initial trajectory straight line, and calculates the orthogonal geometric deviation. Trajectories with geometric deviations not exceeding a preset geometric tolerance threshold are marked as valid debris trajectories, while those exceeding the threshold are marked as background interference trajectories. This module outputs the real image coordinate sequence of the valid trajectories, the corresponding frame number, and the acquisition time to the trajectory reconstruction module 160.

[0137] The modules above exchange data in the order of "data acquisition - target detection - trajectory initiation - trajectory tracking - trajectory verification". The modules maintain a data correspondence through clearly defined frame numbers, target numbers, and trajectory numbers: the frame number identifies the image's temporal location, the target number identifies the detected target within a single frame, and the trajectory number identifies the same fragment across frames. This avoids data association errors caused by substituting different levels of identifiers.

[0138] When the program instructions stored in the memory are executed by the processor, they can implement the steps of the method in any of the above embodiments. Therefore, each functional module of the system corresponds to a method step, but the module name is not limited to using an independent physical device; multiple modules can run on the same processor, or they can be allocated to multiple processing units according to the image processing speed requirements.

[0139] Furthermore, those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0140] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as 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. Figure 1 A means for a process or multiple processes and / or a block diagram (one or more blocks) specifying the functions.

[0141] 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 A process or multiple processes and / or a block diagram (one or more blocks) specifying the functions.

[0142] 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.

[0143] Additionally, any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

Claims

1. A method for detecting explosion debris that integrates physical priors and temporal convolution, characterized in that, include: S1, obtain the image sequence of the explosion field and the coordinates of the blast center; S2, Based on the morphological and motion features of the image frames in the image sequence, candidate fragment targets are selected; S3, Based on the radial relationship between the blast center coordinates and the candidate debris targets, the candidate debris targets in consecutive image frames are associated to establish the candidate debris trajectory; S4. Construct a historical state sequence based on the historical position and motion state of the candidate fragment trajectory, input the historical state sequence into a temporal convolutional network to predict the position of the fragment in subsequent image frames, and associate the candidate fragment target in the subsequent image frames with the predicted position to update the candidate fragment trajectory. S5. Based on the geometric consistency between the candidate fragment trajectories and the coordinates of the blast center, select the effective fragment trajectories from the candidate fragment trajectories.

2. The explosion debris detection method integrating physical priors and temporal convolution as described in claim 1, characterized in that, The morphological features include the stability of dark connected regions in the image frame, and the motion features include the changes of pixels in the image frame relative to the background; The step of filtering candidate fragment targets based on the morphological and motion features of image frames in the image sequence includes: Based on the stability of the dark connected regions, morphological candidate regions are extracted from the image frames; Based on the changes of the pixels relative to the background, the moving pixel region is extracted from the image frame to generate a moving foreground mask; Based on the spatial overlap between the morphology candidate region and the moving foreground mask, the morphology candidate region is filtered to obtain the target to be filtered. The target to be filtered is filtered based on its attribute information to obtain the candidate fragment target.

3. The explosion debris detection method integrating physical priors and temporal convolution as described in claim 2, characterized in that, The maximum stable extreme value region algorithm is used to extract stable dark connected regions in the image frame to form the morphological candidate region; A visual background extraction algorithm is used to extract the moving pixel region in the image frame to generate the moving foreground mask; The spatial overlap is determined as the ratio of the number of pixels belonging to the moving foreground mask within the morphology candidate region to the total number of pixels in the morphology candidate region, and the morphology candidate region with a spatial overlap greater than a preset overlap threshold is determined as the target to be filtered. The attribute information includes pixel area and grayscale. Targets whose pixel area is within a preset pixel area range and meets a preset grayscale condition are identified as candidate fragment targets.

4. The explosion debris detection method integrating physical priors and temporal convolution as described in claim 1, characterized in that, The step of associating candidate debris targets in consecutive image frames based on the radial relationship between the blast center coordinates and the candidate debris targets to establish candidate debris trajectories includes: The candidate fragment targets in the current image frame are identified as targets to be associated. Based on the relative position of the blast center coordinates and the target to be associated in the image coordinate system, determine the radial direction from the blast center coordinates to the target to be associated. Using the radial direction as a matching reference, candidate associated targets whose displacement direction relative to the target to be associated satisfies the preset direction constraint are selected from the candidate fragment targets in the next image frame; From the candidate associated targets, determine the matching target corresponding to the target to be associated, and establish the association relationship between the target to be associated and the matching target; Based on the correlation in adjacent image frames, candidate fragment targets in consecutive image frames are associated to form candidate fragment trajectories.

5. The explosion debris detection method integrating physical priors and temporal convolution as described in claim 4, characterized in that, The step of selecting candidate associated targets from candidate fragment targets in the next image frame whose displacement direction relative to the target to be associated satisfies a preset direction constraint, using the radial direction as a matching reference, includes: Using the position of the target to be associated as a reference and the radial direction as the axis, a fan-shaped physical constraint region is constructed. The search radius of the fan-shaped physical constraint region is determined according to the maximum possible displacement of the fragment, and the angle range of the fan-shaped physical constraint region is determined according to the preset initial tolerance angle. The candidate fragment target located within the fan-shaped physical constraint region in the next image frame is determined as the candidate associated target, wherein the distance between the candidate associated target and the target to be associated does not exceed the maximum possible displacement, and the angle between the displacement direction from the target to be associated to the candidate associated target and the radial direction does not exceed the preset initial tolerance angle.

6. The explosion debris detection method integrating physical priors and temporal convolution as described in claim 4, characterized in that, The step of associating candidate fragment targets in consecutive image frames to form candidate fragment trajectories based on the correlation relationship in adjacent image frames includes: Candidate fragment targets in three consecutive image frames are sequentially associated according to the temporal order of the image sequence; When the association between candidate fragment targets in adjacent image frames is established and the candidate fragment targets in three consecutive image frames pass the preset smoothness test, the candidate fragment targets in the three consecutive image frames are associated to form a candidate fragment trajectory, and the candidate fragment trajectory is activated.

7. The explosion debris detection method integrating physical priors and temporal convolution as described in claim 1, characterized in that, The step of constructing a historical state sequence based on the historical position and motion state of the candidate fragment trajectory, inputting the historical state sequence into a temporal convolutional network to predict the position of the fragment in subsequent image frames, and associating the predicted position with candidate fragment targets in subsequent image frames to update the candidate fragment trajectory includes: Extract the centroid coordinates of the candidate fragment targets from L consecutive historical image frames from the candidate fragment trajectory, where L is a preset number of historical frames; The first-order differential velocity is determined based on the difference in centroid coordinates between adjacent historical image frames. The normalized centroid coordinates and the first-order differential velocity corresponding to each historical image frame are stored in a sliding window in chronological order to construct the historical state sequence. The historical state sequence is input into the temporal convolutional network to obtain the predicted centroid coordinates and error covariance of the fragment in the next image frame. Based on the positional deviation between each candidate fragment target in the next image frame and the predicted centroid coordinates and the error covariance, the Mahalanobis distance of each candidate fragment target relative to the predicted centroid coordinates is determined, and the candidate fragment target with the smallest Mahalanobis distance that is less than a preset threshold determined based on the chi-square distribution is determined as the matching target corresponding to the candidate fragment trajectory. The candidate fragment trajectory is updated according to the matching target.

8. The explosion debris detection method integrating physical priors and temporal convolution as described in claim 7, characterized in that, The temporal convolutional network includes a causal convolutional structure, a dilated convolutional structure, and a residual connection structure, and adopts a three-layer residual block structure with a convolutional kernel size of 3. Updating the candidate fragment trajectory according to the matching target includes: When the matching target is determined, the actual observation state corresponding to the matching target is added to the sliding window to update the historical state sequence, and the matching target is added to the candidate fragment trajectory; When there are no candidate fragment targets with a Mahalanobis distance less than the preset threshold in the next image frame, the predicted state output by the temporal convolutional network at the previous time step is added to the sliding window as the virtual state at the current time step, and the position of the fragment in subsequent image frames is continued to be predicted. When the corresponding candidate fragment target reappears within five image frames and successfully matches the candidate fragment trajectory, the candidate fragment trajectory is resumed; when no corresponding candidate fragment target is matched after five image frames, the candidate fragment trajectory is terminated.

9. The explosion debris detection method integrating physical priors and temporal convolution as described in claim 1, characterized in that, The step of filtering valid debris trajectories from the candidate debris trajectories based on the geometric consistency between the candidate debris trajectories and the coordinates of the blast center, and obtaining the corresponding image coordinate sequence from the valid debris trajectories, includes: From the initial stage of the candidate fragment trajectory, the first M observation points are extracted according to the time order of the image frames, where M is a positive integer not less than 2; The least squares method is used to fit a two-dimensional straight line to the first M observation points to obtain a trajectory straight line that represents the initial motion direction of the candidate fragment trajectory; The orthogonal deviation between the coordinates of the blast center and the trajectory line is defined as the geometric deviation; When the geometric deviation does not exceed the preset geometric tolerance threshold, the candidate fragment trajectory is determined as the effective fragment trajectory; when the geometric deviation exceeds the preset geometric tolerance threshold, the candidate fragment trajectory is determined as background interference trajectory and is eliminated. The real observation points in the effective fragment trajectory are extracted, and the image coordinates of each real observation point are arranged in the time order of the corresponding image frames to obtain the image coordinate sequence.

10. A system for detecting explosive fragments that integrates physical priors and temporal convolution, characterized in that, include: The data acquisition module is used to acquire image sequences of the explosion field and the coordinates of the blast center; The target detection module is used to filter candidate fragment targets based on the morphological and motion features of image frames in the image sequence; The trajectory initiation module is used to associate candidate debris targets in consecutive image frames based on the radial relationship between the blast center coordinates and the candidate debris targets, so as to establish the candidate debris trajectory; The trajectory tracking module is used to construct a historical state sequence based on the historical position and motion state of the candidate fragment trajectory, input the historical state sequence into a temporal convolutional network to predict the position of the fragment in subsequent image frames, and associate the candidate fragment target in the subsequent image frames with the predicted position to update the candidate fragment trajectory. The trajectory verification module is used to filter out valid fragment trajectories from the candidate fragment trajectories based on the geometric consistency between the candidate fragment trajectories and the coordinates of the blast center.

Images