Port container truck positioning method and system based on dynamic MOG and multistage optical flow optimization

Abstract

The invention discloses a port container truck positioning method and system based on a dynamic Gaussian mixture model and multistage optical flow optimization, and belongs to the technical field of computer vision. According to the method, background modeling is carried out on a video sequence by adopting a dynamic Gaussian mixture model to adapt to dynamic environment change, and a container truck area serving as a foreground is accurately extracted; a multi-resolution hierarchical structure is constructed, a multi-stage optical flow optimization strategy is adopted, and a motion vector is calculated and optimized step by step from a low-resolution hierarchy to a high-resolution hierarchy so as to accurately capture large-range rapid movement of a truck; and in combination with the foreground region and the accurate motion vector, accurate positioning and trajectory tracking of the container truck are realized. According to the invention, the dynamic background interference can be effectively overcome, the stable and accurate tracking of the fast moving target is realized, and the precision and robustness of the positioning system in the real operation environment of the port are obviously improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and pattern recognition, and in particular to a port container truck positioning method and system based on dynamic MOG and multi-level optical flow optimization. BACKGROUND

[0002] The port is a key hub in the global logistics chain, and its operation efficiency directly affects the flow speed of international trade. As the main tool for horizontal transportation within the port, the real-time and accurate positioning and tracking of container trucks is a core link to realize the automatic scheduling of the port, improve the operation efficiency, and ensure the safety of production.

[0003] Currently, the technology for positioning container trucks in the port environment mainly relies on video monitoring and image processing. However, the existing technical solutions face many challenges in actual application, resulting in insufficient positioning accuracy and robustness:

[0004] The port background is extremely complex and dynamic: the port is usually by the sea, and the waves and water waves on the sea surface will continuously produce dynamic interference; the frequent movement of large mechanical equipment (such as shore cranes and gantry cranes) can also be misidentified as foreground targets; in addition, changes in light at different times, as well as weather factors such as rain and fog, all pose a serious challenge to the stability of background modeling. Traditional background modeling methods, such as single Gaussian model or frame difference method, are difficult to effectively separate the real target - container truck in such a highly dynamic environment, resulting in high false alarm rate and missed alarm rate.

[0005] The target moves fast and has large displacement: when the container truck drives in the port area, especially in open areas, the speed can be relatively fast, resulting in a large displacement between consecutive frames of video. Traditional, single-level optical flow estimation algorithms (such as the classic Lucas-Kanade method) are prone to exceed their search window limits when dealing with such large displacement movements, resulting in incorrect motion estimation and leading to tracking failure or significant decline in positioning accuracy.

[0006] In summary, the existing methods often perform well under specific, ideal working conditions, but once faced with the superposition of multiple interference factors in the real port operation scenario, their positioning performance will deteriorate sharply, failing to meet the requirements of high precision and high stability of the port automation system.

[0007] Therefore, it is urgent to develop a new target positioning method that can overcome complex and dynamic background interference and effectively handle large displacement movements, in order to achieve all-weather, high-precision, and high-robust positioning and tracking of port container trucks. SUMMARY

[0008] In order to overcome the current problems and defects, the application provides a port container truck positioning method based on dynamic MOG and multi-level optical flow optimization, characterized by comprising the following steps:

[0009] S1. Video acquisition and preprocessing: acquiring a video sequence of a port scene, and preprocessing video frames;

[0010] S2. Dynamic foreground extraction: applying a dynamic Gaussian mixture model (MOG) to process the preprocessed video frames, establishing a dynamic background model, and separating a moving container truck from the video frames as a foreground region;

[0011] S3. Motion vector calculation and repair: for the foreground region, a multi-resolution hierarchical structure is constructed, and a multi-level optical flow method is used to calculate the motion vector of the foreground region; and in the calculation process, the abnormal optical flow area caused by occlusion is repaired;

[0012] S4. Positioning and tracking: combining the foreground region and the repaired motion vector, the accurate position of the container truck is determined, and the motion trajectory is generated.

[0013] Further, in step S2 of dynamic foreground extraction, the learning rate parameter of the Gaussian mixture model is dynamically adjusted to adapt to the moving speed of large equipment such as port cranes, and background false detection is suppressed.

[0014] Further, in step S3 of motion vector calculation and repair, the Lucas-Kanade (Lucas-Kanade) optical flow algorithm is used to calculate the motion vector in each level of the multi-resolution hierarchical structure.

[0015] Further, the repair of the abnormal optical flow area comprises:

[0016] Through direction consistency detection, the motion vector deviating from the main motion direction of the foreground region by more than a preset angle threshold is marked as an abnormal vector; and

[0017] The abnormal vector is replaced by using neighborhood interpolation or prediction interpolation based on historical trajectory.

[0018] Further, step S4 of positioning and tracking comprises:

[0019] Spatial clustering analysis is performed on the repaired motion vectors in the foreground region;

[0020] After removing the outlier vectors deviating from the cluster center, the weighted average value of the remaining vectors is calculated to determine the overall motion of the container truck;

[0021] The truck position is updated according to the overall motion, and the motion trajectory is generated.

[0022] Furthermore, the method also includes a dynamic optimization step for computing resources, which includes:

[0023] The computing resources allocated to subsequent processing steps are dynamically adjusted based on the area of ​​the foreground region or the system processing delay.

[0024] Furthermore, the dynamic optimization step for computing resources further includes at least one of the following:

[0025] When the area of ​​the foreground region is less than a first threshold, the number of threads used for parallel computing is reduced.

[0026] When the system processing delay continues to exceed the second threshold, the resolution of the video frame to be processed is reduced.

[0027] The present invention also provides a port container truck positioning system, comprising:

[0028] Video acquisition and preprocessing module: used to acquire video sequences of port scenes and preprocess video frames;

[0029] Dynamic foreground extraction module: used to apply the Dynamic Gaussian Mixture Model (MOG) to process the preprocessed video frames, establish a dynamic background model, and separate the moving container truck from the video frames as the foreground region;

[0030] Motion vector calculation and repair module: This module is used to calculate the motion vector of the foreground region by constructing a multi-resolution hierarchical structure and using a multi-level optical flow method; and is configured to repair areas with abnormal optical flow caused by occlusion during the calculation process.

[0031] Positioning and tracking module: used to determine the precise location of the container truck and generate its trajectory by combining the foreground area and the repaired motion vector.

[0032] Furthermore, the system also includes a dynamic optimization module for computing resources, which is configured as follows:

[0033] The computing resources allocated to subsequent processing steps are dynamically adjusted based on the area of ​​the foreground region or the system processing delay.

[0034] Furthermore, the computing resource dynamic optimization module is further configured to perform at least one of the following operations:

[0035] When the area of ​​the foreground region is less than a first threshold, the number of threads used for parallel computing is reduced.

[0036] When the system processing delay continues to exceed the second threshold, the resolution of the video frame to be processed is reduced.

[0037] Beneficial effects:

[0038] This invention employs a Dynamic Gaussian Mixture Model (MOG) for background modeling. This model can adaptively learn and update the background, effectively distinguishing moving container trucks from dynamic background interference such as sea waves, changes in lighting, and the movement of large equipment. This fundamentally solves the problem of traditional methods easily misjudging dynamic backgrounds as foreground in complex port environments, significantly improving the accuracy and completeness of foreground target (container truck) extraction.

[0039] This invention introduces a multi-level optical flow optimization strategy based on an image pyramid. By calculating the initial optical flow starting from the low-resolution image at the top of the pyramid and progressively passing the results to high-resolution image layers for optimization, it can effectively capture and calculate the large-scale displacement of container trucks caused by rapid movement. This method overcomes the limitations of traditional single-level optical flow algorithms, obtaining a more accurate and robust motion vector field, thereby significantly improving the final positioning accuracy.

[0040] This invention combines precise foreground extraction with reliable motion analysis to form a complete positioning scheme. This scheme not only boasts high positioning accuracy but also meets the real-time requirements of port operations due to the high computational efficiency of the MOG algorithm and multi-level optical flow method. The collaborative work of the system's various modules ensures that the positioning results maintain high stability and continuity in the face of the complex and ever-changing real-world environment of ports, demonstrating strong robustness. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of the positioning method of the present invention.

[0043] Figure 2 This is a schematic diagram of the MOG fusion optical flow method of the present invention.

[0044] Figure 3 This is a system architecture diagram of an embodiment of the present invention. Detailed Implementation

[0045] The present application will be described below with reference to specific embodiments:

[0046] Example 1:

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and a specific embodiment. It should be understood that the specific embodiment described herein is merely illustrative and not intended to limit the scope of protection of this invention.

[0048] This embodiment provides a port container truck positioning method and system based on dynamic Gaussian mixture model and multi-level optical flow optimization.

[0049] The positioning system hardware in this embodiment can be based on the port's existing video surveillance network. For example... Figure 1 As shown, the system can be deployed on one or more servers with sufficient computing power (such as GPUs for parallel computing acceleration) to process video streams. On the software side, the system includes the following core functional units:

[0050] Video acquisition unit: calls high-definition surveillance cameras deployed at key locations in the port operation area (such as the front of the wharf and the yard passage) to acquire real-time video streams with a resolution of, for example, 1920x1080 and a frame rate of, for example, 25fps.

[0051] Image preprocessing unit: processes each received frame of color video image to prepare for subsequent steps.

[0052] Background modeling and foreground extraction unit: The core function is to separate the moving container truck.

[0053] Motion Analysis Unit: Its core function is to calculate the precise motion information of the truck.

[0054] Positioning and tracking unit: Its core function is to integrate information and output the final position and trajectory.

[0055] Based on the above methods, such as Figure 2 As shown, this embodiment provides the specific steps of a port container truck positioning method based on a dynamic Gaussian mixture model and multi-level optical flow optimization:

[0056] Step 1: Video Acquisition and Preprocessing

[0057] The system continuously acquires real-time video streams of the port's operational area through a video acquisition unit. Each frame of the video stream is transmitted to an image preprocessing unit. In this unit, to reduce the complexity of subsequent calculations and eliminate interference from color information on optical flow and background modeling, each frame of the RGB color image is converted to grayscale, transforming it into an 8-bit single-channel grayscale image. This step simplifies the data while preserving the image's brightness, contour, and texture information, which is crucial for target recognition and motion analysis.

[0058] Step 2: Dynamic Background Modeling and Foreground Extraction

[0059] The preprocessed grayscale image sequence is fed into the background modeling and foreground extraction unit. For example... Figure 3 As shown, this unit uses the Dynamic Gaussian Mixture Model (MOG) algorithm to build a background model that can be adaptively updated.

[0060] Unlike static methods, MOG maintains a mixture model for each pixel in the image, consisting of K Gaussian distributions (K can be 3 to 5 in this embodiment). When a new video frame is input, the current pixel value is matched against the K Gaussian distributions at that point:

[0061] Matching and Updating: A match is considered successful if the difference between the current pixel value and the mean of a Gaussian distribution is within a certain range (e.g., 2.5 standard deviations). The weights of successfully matched Gaussian distributions are increased, while the weights of other distributions are decreased. Simultaneously, the mean and variance of the matched distribution are updated based on the current pixel value. This mechanism allows the model to learn slow changes in the background, such as gradual changes in lighting.

[0062] Background / Foreground Determination: K Gaussian distributions are sorted by weight. The distributions at the top of the list, representing the scene most of the time, are considered background components. If the current pixel value cannot match any background component, the pixel is determined to be foreground.

[0063] Dynamic Adaptability: A key advantage of the MOG algorithm is its ability to automatically adjust the number of Gaussian distributions (K value) for each pixel as needed, and it possesses a learning rate parameter that controls how quickly the background model adapts to scene changes. In a port environment, facing the periodic fluctuations of waves or the intermittent movement of gantry cranes, MOG2 can learn to gradually incorporate these dynamic elements into the background model, thereby ensuring that only non-background, independent moving targets (i.e., container trucks) are stably extracted as foreground elements.

[0064] After this step, the system outputs a binary foreground mask image, where the white areas represent the detected container trucks and the black areas represent the background.

[0065] Step 3: Multi-level optical flow analysis and motion vector calculation

[0066] Constructing a multi-resolution hierarchical structure: First, for and A series of Gaussian smoothing and downsampling operations are performed to construct an image hierarchy with multiple resolutions. The bottom layer of this structure is the original image, and each layer above it is obtained by downsampling half of the image below it. The top layer of this structure has the lowest resolution, but the macroscopic motion of the target is most significant.

[0067] Stepwise optical flow calculation and optimization:

[0068] The calculation process begins at the top layer (the lowest resolution layer). Due to the low image resolution, the truck's large movements are represented as small displacements at this layer. At this point, the system uses the Lucas-Kanade (LK) optical flow method to calculate an initial optical flow field. .

[0069] Then, the optical flow field Magnify it by two times and use it as the initial displacement for calculating the optical flow of the next layer (layer L-2). Specifically, when calculating the optical flow of the image in layer L-2, first use... compensate The L-2 layer image, and then the compensated image with Calculate the optical flow residual between the L-2 layer images The final optical flow result for this layer is the sum of the initial displacement and the residual: .

[0070] Repeat this process of "transfer-compensation-calculation of residuals" step by step down until the optical flow field of the original image layer (layer 0) is calculated.

[0071] This multi-level strategy decomposes a difficult large-displacement optical flow problem into a series of simple small-displacement optical flow problems for solving, thereby enabling accurate calculation of the pixel-level motion vector of the container truck between two frames.

[0072] Step 4: Localization and Tracking

[0073] Finally, the positioning and tracking unit integrates the results of the second and third steps to complete the final positioning.

[0074] Extracting the motion vector of the target region: The system combines the foreground mask image obtained in the second step with the full-image optical flow field obtained in the third step. Perform an AND operation. This way, only motion vectors belonging to the container truck area are preserved.

[0075] Calculating overall motion and position: To obtain the overall motion of the truck as a rigid body, the system performs "region aggregation and filtering" on all valid motion vectors within the foreground region:

[0076] Spatial clustering: Clustering algorithms such as DBSCAN are used to spatially cluster all motion vectors within the region, grouping vectors with similar motion patterns into a main motion cluster.

[0077] Outlier removal: Calculate the center (mean) and standard deviation of the main motion cluster. Remove all outlier vectors that deviate from the cluster center by more than 2 standard deviations. This step effectively filters out sporadic erroneous vectors caused by reflections or local occlusion.

[0078] Weighted average: Finally, the remaining effective vectors after the above cleaning steps are weighted and averaged to obtain a final motion vector that can accurately represent the overall translation of the truck.

[0079] Position Update and Trajectory Generation: The center position coordinates of the truck from the previous moment are updated based on this final motion vector to obtain the precise position coordinates of the truck at the current moment. By continuously recording the calculated position coordinates for each frame, the real-time motion trajectory of the container truck within the port area is formed. Simultaneously, the system can also analyze the truck's real-time speed and direction of travel, among other status information, based on the magnitude and direction of the motion vector.

[0080] Through the above steps, this embodiment can overcome the interference of complex dynamic backgrounds in ports and achieve stable and accurate positioning and tracking of fast-moving container trucks.

[0081] Example 2:

[0082] This embodiment provides a positioning process for a typical container truck positioning scenario, as described below:

[0083] In the port's quay crane operation area, 2560×1440 resolution industrial cameras (60fps) are deployed, with calibration parameters of 4.2mm / pixel. Currently, the target area for trucks occupies approximately 20%, with sporadic container obstructions, and the ambient light intensity is 800 lux.

[0084] Implementation steps:

[0085] MOG Dynamic Foreground Segmentation:

[0086] Initialize the Gaussian mixture model, set the Gaussian distribution number K=3, learning rate α=0.01, and background judgment threshold T=0.7; model 5 consecutive frames of images, separate the dynamic foreground (truck) from the static background (dock ground), and output a binarized mask, with the dynamic region covering the pixel coordinate range (320,240)-(1440,810).

[0087] Gaussian mixture model probability density function:

[0088]

[0089] Where η is a Gaussian distribution, α = 0.005, and the control parameter update rate is:

[0090]

[0091] =1 indicates a successful match, otherwise it is 0.

[0092] Positioning distance calculation:

[0093] The Lucas-Kanade optical flow algorithm is applied to the region within the mask, with a window size of 15×15 and a maximum number of iterations of 20. An abnormal displacement direction (deviation of 25°) is detected near the coordinates (1000, 600) due to container obstruction. The median of the effective displacement in the 3×3 neighborhood (Δx=120 pixels, Δy=5 pixels) is used to replace the abnormal value.

[0094] Taylor expansion with gray-level invariant constraints:

[0095]

[0096] Minimize the residual within the window:

[0097]

[0098] The solution is:

[0099]

[0100] Convert pixel displacement to actual distance:

[0101]

[0102] DBSCAN clustering (ε=15 pixels, min_samples=6) was used to remove 8 outliers, and the average deviation was calculated to be 0.50m.

[0103] Dynamic resource allocation:

[0104] The dynamic region accounts for 25%, activating the corresponding GPU thread block (128×128 grid), with a processing latency of 14ms / frame.

[0105] Example 3:

[0106] This embodiment presents a truck positioning process in a low-light, multi-vehicle convergence scenario at night. The scenario is described as follows:

[0107] In a nighttime operation scenario, the intersection of two trucks resulted in the target area occupying 45% of the space and having a 60% obstruction rate, prompting the activation of infrared supplementary lighting.

[0108] Implementation steps:

[0109] MOG parameter dynamic adjustment:

[0110] The system detected that the foreground area ratio (45%) and rate of change both exceeded the preset dynamic threshold due to the intersection of two vehicles, indicating that the scene had entered a high-dynamic mode. To adapt to this change, the system automatically increased the learning rate α of the background model update from the usual 0.01 to 0.05, while simultaneously reducing the background judgment threshold T to 0.5. Through this "variable learning rate" adjustment, the model can more quickly absorb transient and large-scale changes such as headlight halos at night into the background, significantly enhancing the sensitivity of dynamic foreground recognition in complex nighttime scenes and effectively separating interference from dock lights.

[0111] Optical flow restoration and enhancement:

[0112] Continuous occlusion was detected in the coordinate region (600, 400) - (1200, 720), and three-frame linear extrapolation compensation was enabled:

[0113] The optical flow vector sequence of the first 3 frames is (2.1, 1.7) → (2.3, 1.8) → (2.5, 1.9); the predicted vector of the current frame is (2.7, 2.0), and the error between the vector and the actual repaired vector is <0.2 pixels.

[0114] Predicting the current frame from a three-frame sequence (v1, v2, v3):

[0115] :

[0116] In the case =(2.5,1.9)+(0.2,0.1)=(2.7,2.0).

[0117] Adaptive resolution switching:

[0118] If the GPU latency is greater than 25ms for 5 consecutive frames, the resolution will be downgraded to 1280×720 and the optical flow calculation window will be reduced to 10×10. After the downgrade, the latency will be reduced to 18ms / frame, and 1080P output will be restored by bilinear interpolation with a position error increase of only 0.3%.

[0119] Pixel value calculation during resolution restoration:

[0120]

[0121] Where a and b are the sub-pixel offsets.

[0122] Example 4:

[0123] This embodiment presents a truck positioning process in an extreme reflective multi-target scenario. The scenario is described as follows:

[0124] The metal container surface was highly reflective at noon (illuminance > 1500 lux), and the three trucks traveling side by side caused the target area to occupy 65% ​​of the area, resulting in fragmented obscured areas.

[0125] Implementation steps:

[0126] Multi-level optical flow restoration:

[0127] After detecting fragmented occlusion areas caused by three vehicles driving side-by-side, the system initiates a multi-level repair process:

[0128] Level 1 Repair: Outlier Marking and Neighborhood Interpolation. First, the system statistically analyzes the main motion direction of all motion vectors within the foreground region of the current frame. Then, it performs direction consistency detection, marking vectors deviating from the main direction by more than 15° as isolated outliers. For these outliers, the median of the valid motion vectors within a 5×5 neighborhood is used for replacement. This step repairs approximately 15% of the isolated outlier data.

[0129] Second-level repair: Predictive compensation based on historical trajectories. For continuous occlusion regions with no valid values ​​in the neighborhood (e.g., coordinates (200-400, 300-500)), the system retrieves the valid motion vector data of this region from the previous 5 frames, fits a quadratic curve based on this time series, and uses this curve to predict the motion vector of the occluded region in the current frame. Using this method, the maximum prediction error is controlled within 0.5 pixels, compressing the data loss rate of over 30% in the comparison file CN200610116504.5 to below 5%.

[0130] Extreme resource regulation:

[0131] With a dynamic region coverage of >60%, multi-GPU parallel computing (2 NVIDIA T4 GPUs) is enabled, and latency is maintained at 22ms / frame.

[0132] Theoretical speedup ratio (Amdahl's Law):

[0133]

[0134] In this case, the dual T4 GPUs stabilized the latency at 22ms / frame.

[0135] The general resource allocation model in the above embodiments is determined by the following formula:

[0136] GPU-threaded grid computing:

[0137] Grid dimension =

[0138] For example, in Case 1, 25% of the dynamic region corresponds to the number of threads:

[0139] .

[0140] The above description is only one specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto.

[0141] It should be noted that the specific parameters set in this embodiment, such as video resolution, frame rate, number of Gaussian components, and number of image pyramid layers, are merely exemplary values ​​used to clearly illustrate the technical solution of this invention. Those skilled in the art can adjust them according to actual application scenarios and hardware conditions. Similarly, the specific algorithm used (such as the Lucas-Kanade method) can also be replaced by other algorithms with the same function and that can realize the core idea of ​​this invention.

[0142] Any person skilled in the art can make various corresponding changes or modifications without departing from the core ideas and principles of the present invention, and such changes or modifications should all fall within the protection scope of the appended claims. Therefore, the patent protection scope of the present invention should be determined by the appended claims.

Claims

1. A port container truck positioning method based on dynamic MOG and multi-level optical flow optimization, characterized in that, Includes the following steps: S1. Video Acquisition and Preprocessing: Acquire video sequences of the port scene and preprocess the video frames; S2. Dynamic Foreground Extraction: Apply the Dynamic Gaussian Mixture Model (MOG) to process the preprocessed video frames, establish a dynamic background model, and separate the moving container truck from the video frames as the foreground region; S3. Motion Vector Calculation and Repair: For the foreground region, the motion vector of the foreground region is calculated by constructing a multi-resolution hierarchical structure and using a multi-level optical flow method; And during the calculation process, areas with abnormal optical flow caused by occlusion are repaired; S4. Positioning and Tracking: Combining the foreground area and the repaired motion vector, determine the precise position of the container truck and generate its motion trajectory.

2. The method according to claim 1, characterized in that, In step S2, dynamic foreground extraction, the learning rate parameter of the Gaussian mixture model is dynamically adjusted to adapt to the moving speed of large equipment such as port cranes and suppress false background detection.

3. The method according to claim 1, characterized in that, In step S3, motion vector calculation and repair, the Lucas-Kanade optical flow algorithm is used to calculate motion vectors in each level of the multi-resolution hierarchical structure.

4. The method according to claim 1 or 3, characterized in that, The repair of the optical flow anomaly region includes: Through orientation consistency detection, motion vectors whose deviation from the main motion direction of the foreground region exceeds a preset angle threshold are marked as abnormal vectors; and The abnormal vector is replaced using neighborhood interpolation or prediction interpolation based on historical trajectories.

5. The method according to claim 1, characterized in that, Step S4, positioning and tracking, includes: Spatial clustering analysis is performed on the repaired motion vectors within the foreground region; After removing outlier vectors that deviate from the cluster center, the weighted average of the remaining vectors is calculated to determine the overall motion of the container trucks; The truck position is updated based on the overall motion, and a motion trajectory is generated.

6. The method according to claim 1, characterized in that, The method also includes a dynamic optimization step for computing resources, which includes: The computing resources allocated to subsequent processing steps are dynamically adjusted based on the area of ​​the foreground region or the system processing delay.

7. The method according to claim 6, characterized in that, The dynamic optimization step for computing resources further includes at least one of the following: When the area of ​​the foreground region is less than a first threshold, the number of threads used for parallel computing is reduced. When the system processing delay continues to exceed the second threshold, the resolution of the video frame to be processed is reduced.

8. A port container truck positioning system, characterized in that, include: Video acquisition and preprocessing module: used to acquire video sequences of port scenes and preprocess video frames; Dynamic foreground extraction module: used to apply the Dynamic Gaussian Mixture Model (MOG) to process the preprocessed video frames, establish a dynamic background model, and separate the moving container truck from the video frames as the foreground region; Motion vector calculation and repair module: This module is used to calculate the motion vector of the foreground region by constructing a multi-resolution hierarchical structure and using a multi-level optical flow method. It is configured to repair areas of abnormal optical flow caused by occlusion during the calculation process; Positioning and tracking module: used to determine the precise location of the container truck and generate its trajectory by combining the foreground area and the repaired motion vector.

9. The system according to claim 8, characterized in that, The system also includes a dynamic optimization module for computing resources, which is configured as follows: The computing resources allocated to subsequent processing steps are dynamically adjusted based on the area of ​​the foreground region or the system processing delay.

10. The system according to claim 9, characterized in that, The computing resource dynamic optimization module is further configured to perform at least one of the following operations: When the area of ​​the foreground region is less than a first threshold, the number of threads used for parallel computing is reduced. When the system processing delay continues to exceed the second threshold, the resolution of the video frame to be processed is reduced.

Images