Coal mine drilling rod continuous tracking method based on geometric correlation and angle threshold

Abstract

The present application relates to a coal mine drilling rod continuous tracking method based on geometric correlation and angle threshold in the field of artificial intelligence and computer vision technology, a tracking framework suitable for rotating targets is constructed based on ByteTrack algorithm, target detection is carried out on video data through ByteTrack algorithm; the detection results are divided according to the confidence score; the high-score detection frame is matched with the prior tracking frame through IOU; the prior tracking frame that is not matched successfully in the initial matching is matched with the low-score detection frame through IOU; the unactivated tracking frame is matched with the high-score detection frame that is not matched successfully through IOU; anchor point rescue matching, the high-score detection frame that is not matched successfully is matched with the prior tracking frame that is not matched successfully; reservation and new creation. The present application realizes more robust, continuous and reliable multi-target tracking under complex downhole conditions, maintains target identity consistency and trajectory coherence, provides more reliable data support for drilling footage calculation, and significantly improves the stability and reliability of the data relied on footage statistics.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and computer vision technology, and provides a method for continuous tracking of drill rods in coal mine drilling based on geometric correlation and angle threshold. Specifically, it involves video target detection and mineral geological exploration technology using high technology, and is particularly suitable for drilling rig target recognition and detection in underground coal mine drilling operation environment. Background Technology

[0002] In underground drilling operations in coal mines, accurate acquisition of borehole footage relies on the stable identification of the drilling rig target. In real-world scenarios, factors such as low light conditions, image blur, and frequent occlusion can negatively impact drilling rig identification. Based on rotating target detection, this identification is prone to intermittent detection, momentary missed detections, and even target disappearance, thus weakening the stability and reliability of footage statistics. Existing technologies often rely on drill rod counting techniques based on rotating target detection, such as peak counting based on drill rod length. While these technologies can identify the drilling rig target in a single frame, they are essentially frame-by-frame detection mechanisms, lacking the continuity constraint of time. In underground operating scenarios with low light, motion blur, and frequent occlusion, the detection results still frequently exhibit discontinuous detection, brief missed detections, or even target loss, affecting the reliability of footage statistics. Therefore, achieving continuous perception of the drilling rig target in complex underground environments is a key issue in improving the intelligence level of operation monitoring. Summary of the Invention

[0003] To address the above issues, this invention provides a continuous tracking method for drill rods in coal mine drilling based on geometric correlation and angle thresholds. This method aims to solve problems such as discontinuous drilling rig identification and detection, brief missed detections, and even target loss in existing underground coal mine drilling operations, which affect the reliability of drilling footage statistics.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A continuous tracking method for coal mine drilling rods based on geometric correlation and angle thresholds is proposed. This method constructs a tracking framework suitable for rotating targets based on the ByteTrack algorithm, and includes the following steps: 1. Object detection: Target detection is performed on video data using the ByteTrack algorithm; 2. Classify the detection results: Classify the detection results according to the confidence score, and set two thresholds s1 and s2 to classify the detection result detection box O as a high-scoring detection box O. h and low-resolution detection boxes O l Detection boxes with excessively low scores are discarded. High-scoring detection boxes correspond to more accurate detection targets, while low-scoring detection boxes contain falsely detected or inaccurately detected targets. The posterior tracking bounding box K obtained from the previous frame of the video data (tracking trajectory) tracked Based on whether it is activated, it is divided into activation tracking box K. active Inactive tracking box K uncomfirmed Place the posterior tracking box that is in a lost state and has not been lost more than 30 times into the lost tracking box K. loss In the middle, activate the tracking box K active With the lost tracking frame K loss The prior tracking box K of the current frame is obtained through Kalman filter prediction. all ; III. Initial matching: high-resolution detection bounding boxes O h With prior tracking box K all The IOU is calculated to obtain the IOU cost matrix C1. The Hungarian algorithm is then used to perform association matching based on C1 and the adaptive matching threshold h1. After matching, the matched prior tracking boxes K are obtained. m With high-resolution detection box O h-m Unmatched prior tracking box K r High-resolution detection bounding boxes O that did not match successfully h-r， The high-resolution detection bounding box that was successfully matched O h-m The prior tracking box K is updated using Kalman filtering. m The new posterior tracking box K is then obtained. m ; IV. Secondary matching: Prior tracking boxes K that did not match successfully in the initial matching. r With low-resolution detection box O l Perform IOU matching, calculate the IOU cost matrix C2 between the two, and use the Hungarian algorithm to perform association matching based on C2 and the adaptive matching threshold h2 to obtain the matched prior tracking boxes (trajectories) K. rm With low-resolution detection box O l-m Unmatched prior tracking box K rr Low-scoring detection boxes O that did not match successfully l-r The low-scoring detection box O after successful matching l-m The prior tracking box K is updated using Kalman filtering. rm The new posterior tracking box K is then obtained. rm ; 5. Three matches, tracking box K not activated. uncomfirmed High-resolution detection boxes O that did not match successfully h-r Perform IOU matching and calculate the inactive tracking box K. uncomfirmed High-resolution detection boxes O that did not match successfully h-r The IOU cost matrix C3 is used to perform association matching between C3 and the adaptive matching threshold h3, resulting in the matched inactive tracking boxes K. uncomfirmed-mWith high-resolution detection box O h-rm Unmatched inactive tracking box K uncomfirmed-r High-resolution detection boxes that did not match successfully h-rr The high-scoring detection boxes that match successfully will be O h-rm Update the inactive tracking box K using Kalman filtering uncomfirmed-m The activated tracking box K is then obtained. uncomfirmed-m (Also called the posterior tracking box K) uncomfirmed-m ); VI. Anchor point rescue matching: High-resolution detection boxes O that failed to match in step five. h-rr The prior tracking box K that did not match successfully in step four rr Matching is performed. Since this is multi-class, multi-object tracking, the prerequisite for this step is matching within the same class to obtain the matched prior tracking bounding boxes (trajectories) K. rrm With high-resolution detection box O h-rrm Unmatched prior tracking box K rrr High-resolution detection bounding boxes O that did not match successfully h-rrr The high-scoring detection box O after successful matching h-rrm The prior tracking box K is updated using Kalman filtering. rrm The new posterior tracking box K is then obtained. rrm ; 7. Retention and Creation: Prior tracking boxes that failed to match in step six are retained by the ByteTrack algorithm for 30 frames and subjected to IOU matching upon their reappearance. Prior tracking boxes with more than 30 consecutive failed matches are discarded, while those with no more than 30 consecutive failed matches are directly used as posterior tracking boxes. High-resolution detection boxes that failed to match in step six are recreated by the ByteTrack algorithm as new tracking boxes K. n Newly created tracking boxes are directly used as posterior tracking boxes; retained prior tracking boxes are placed into the lost tracking box K. loss In the middle, the posterior tracking box K m Posterior tracking box K rm Activate tracking box K uncomfirmed-m Posterior tracking box K rrm Create a new tracking box K n These are placed in the posterior tracking bounding box K. tracked In the middle, the posterior tracking box K tracked The active tracking box is the one used to display on the current video frame.

[0006] 8. Repeat the above steps until there is no more video data.

[0007] Furthermore, the matching degree determination in step six is ​​as follows: The number of key points of the unmatched high-resolution detection boxes located within the range circle of the unmatched prior tracking boxes in step four is obtained, and this number is multiplied by... The score is used as the matching criterion, with higher scores indicating higher matching accuracy. Here, d represents the distance between the center points of the high-scoring detection boxes that failed to match in step five and the prior tracking boxes that failed to match in step four, and the decay rate is controlled by the parameter 1 / 20. Traditional IOU cost matrices often lead to target loss due to low or even zero IOU values ​​between detection and tracking boxes when addressing the aforementioned problems. This method uses a score-based mechanism instead of the traditional IOU cost matrix as the matching criterion, allowing for more efficient utilization of high-scoring detection boxes. Specific matching rules effectively solve the problems of target occlusion and nonlinear motion.

[0008] Furthermore, the key points are the four corner points and the center point of the high-resolution detection frame.

[0009] Furthermore, the radius of the prior tracking box is the larger of its width and height, and the center point of the prior tracking box is the center of the circle. This effectively increases the matching range of the prior tracking box.

[0010] Furthermore, the IOU matching method in steps three, four, and five involves calculating the corresponding IOU. angle The cost matrix is ​​used, and the Hungarian algorithm is employed to perform association matching based on the cost matrix and an adaptive matching threshold. The IOU... angle The cost matrix is ​​as follows:

[0011] Among them, |θ detect -θ track | represents the difference in angle between the corresponding detection box and the tracking box; α is a parameter that reflects the angle difference between the detection box and the tracking box. The larger the angle difference (the less similar the two boxes are in terms of angle), the smaller the value of α, and the smaller the angle difference (the more similar the two boxes are in terms of angle), the larger the value of α.

[0012] IOU represents a matrix that calculates the intersection-union ratio (iou) for each of the detection boxes (detect1, detect2, detect3...) and the tracking boxes (track1, track2, track3...). (For example, the horizontal axis represents the detection boxes and the vertical axis represents the tracking boxes.)

[0013] IOU angle The cost matrix is ​​obtained by multiplying the IOU matrix by the α corresponding to the element in the matrix.

[0014] Cost matrix: a matrix of all 1s minus IOUangle The cost matrix is ​​obtained.

[0015] The formula for the adaptive matching threshold is as follows: .

[0016] Here, Fixed_threshold is the original matching threshold set manually by bytetrack. Auto_Threshold dynamically adjusts the original Fixed_threshold using an α value (obtained after filtering) in the current frame. The matching threshold is used to filter cells in the cost matrix Cost, checking if their values ​​are less than the threshold. Only cells with values ​​less than the threshold (detection boxes and tracking boxes) have a chance of being paired. Tracking ensures that the tracking target ID remains unchanged through correct pairing.

[0017] To reflect the impact of target rotation angle on matching in the cost matrix, |θdetect-θtrack|, i.e., the detection box θ detect With tracking box θ track The angle difference. Since the above involves multi-target, multi-class detection and tracking, the cost value between different classes is directly set to 1. The angle difference between the detection and tracking boxes of different classes is no longer used in the cost matrix to prevent changes in the tracking box category. Because multiplying the cost matrix by α effectively reduces its size, the corresponding thresholds h1, h2, and h3 in the matching stage also change accordingly. Here, it is determined by taking the value from |θdetect-θtrack| that is actually used in the cost matrix and where the distance between the center points of the detection and tracking boxes is less than a certain distance (half the larger of the detection box's length and width plus half the larger of the tracking box's length and width). The larger of the two is used to modify the adaptive matching threshold. Through the above optimizations, the inherent limitations of traditional ByteTrack in rotating target scenarios are overcome, enabling the tracking system to not only maintain high accuracy under normal conditions, but also achieve more robust, stable and reliable rotating multi-target tracking in extremely complex environments such as underground coal mines.

[0018] The beneficial effects of this invention are as follows: To address the shortcomings of existing technologies that rely solely on single-frame rotating target detection, which cannot fully utilize the temporal information during underground coal mine operations and cannot maintain continuous target observation during occlusion or temporary missed detections, lacking temporal continuity constraints, this invention addresses the continuous perception requirements of rotating targets in underground coal mine drilling rig operations. In low-light, motion-blurred, and frequently occluded underground operation scenarios, detection results often exhibit discontinuous detection, temporary missed detections, or even target loss, thus affecting the reliability of footage statistics. This invention introduces a robust rotating target tracking method to meet the continuous perception needs of rotating targets in underground coal mine drilling rig operations. By ensuring the target's identity remains unchanged and its trajectory is continuous throughout the time series, and considering the characteristics of nonlinear motion and frequent occlusion in underground scenarios, the ByteTrack matching mechanism is improved. Firstly, new association indicators are constructed by introducing the length, width, center point, and geometric relationships of the rotating bounding box, improving the utilization rate of high-resolution detection boxes. Secondly, a new cost matrix is ​​designed to take into account the influence of the angle difference between the detection box and the tracking box, and an Angle-aware Adaptive Threshold (AAT) is added to dynamically adjust the threshold of the association stage based on the angle difference between the detection box and the tracking box. Overall, this method overcomes the limitations of traditional ByteTrack in rotating target scenarios, achieving more robust, continuous, and reliable multi-target tracking under complex downhole conditions. It maintains target identity consistency and trajectory coherence, thus providing more reliable data support for borehole footage calculation and significantly improving the stability and reliability of the data upon which footage statistics depend. Attached Figure Description

[0019] Figure 1 This is a framework diagram of the continuous tracking method of the present invention; Figure 2 This is a comparison diagram of the tracking box matching range before and after the improvement in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be clearly described in conjunction with the accompanying drawings and specific embodiments. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art based on the embodiments of the present invention without inventive effort to obtain all other embodiments should be included within the scope of protection of the present invention.

[0021] Example To address the problems of discontinuous detection results, brief missed detections, and even target loss in underground drill pipe counting in coal mines, which affect the reliability of drilling footage statistics, this invention proposes a continuous tracking method for coal mine drill pipes based on geometric correlation and angle thresholds.

[0022] This method is essentially an improved version of ByteTrack for rotating target tracking. By ensuring the target's identity remains unchanged and its trajectory is continuous throughout the time series, it significantly improves the stability and reliability of the data upon which drilling footage statistics depend. This provides stable detection results for subsequent drill pipe counting. Specifically, this invention enhances the original ByteTrack's two-stage association mechanism in two aspects for rotating targets. First, after processing inactive tracking boxes, a geometric anchor point rescue matching step for rotating targets is introduced. Tracking boxes that failed to match in the previous stage are re-associated with unmatched high-resolution detection boxes, and a geometric similarity metric based on corner distance and center distance is used to further improve the association capability for targets with significant deformation and pose changes. Second, in the first three matches, the angle difference between the detection box and the tracking box is included in the cost matrix, enabling the matching process to explicitly perceive changes in the target's direction. An adaptive threshold strategy based on the angle difference is designed to dynamically adjust the matching conditions. This improves the association stability and robustness in rotating and jittery scenarios.

[0023] This method for continuous tracking of drill rods in coal mine drilling, based on geometric correlation and angle thresholds, constructs a tracking framework suitable for rotating targets based on the ByteTrack algorithm, such as... Figure 1 As shown, the specific steps are as follows: 1. Object detection: Target detection is performed on video data using the ByteTrack algorithm; 2. Classify the detection results: Based on the confidence score, classify the detection results by setting two thresholds, s1 and s2, and classify the detection result bounding box O as the high-scoring detection box O. h and low-resolution detection boxes O l Detection boxes with excessively low scores are discarded. High-scoring detection boxes correspond to more accurate detection targets, while low-scoring detection boxes contain false detections or targets that are not detected accurately. The posterior tracking bounding box K obtained from the previous frame of the video data (tracking trajectory) tracked Based on whether it is activated, it is divided into activation tracking box K. active Inactive tracking box K uncomfirmed Place the posterior tracking box that is in a lost state and has not been lost more than 30 times into the lost tracking box K. loss In the middle, activate the tracking box K active With the lost tracking frame K loss The prior tracking box K of the current frame is obtained through Kalman filter prediction. all ; III. Initial matching: high-resolution detection bounding boxes O h With prior tracking box K all The IOU is calculated to obtain the IOU cost matrix C1. The Hungarian algorithm is then used to perform association matching based on C1 and the adaptive matching threshold h1. After matching, the matched prior tracking boxes K are obtained.m With high-resolution detection box O h-m Unmatched prior tracking box K r High-resolution detection boxes O that did not match successfully h-r， The high-resolution detection bounding box that was successfully matched O h-m The prior tracking box K is updated using Kalman filtering. m The new posterior tracking box K is then obtained. m ; IV. Secondary matching, matching, prior tracking boxes K that did not match successfully in the initial matching. r With low-resolution detection box O l Perform IOU matching, calculate the IOU cost matrix C2 between the two, and use the Hungarian algorithm to perform association matching based on C2 and the adaptive matching threshold h2 to obtain the matched prior tracking boxes (trajectories) K. rm With low-resolution detection box O l-m Unmatched prior tracking boxes K rr Low-scoring detection boxes that failed to match l-r The low-scoring detection box O after successful matching l-m The prior tracking box K is updated using Kalman filtering. rm The new posterior tracking box K is then obtained. rm ; 5. Three matches, tracking box K not activated. uncomfirmed High-resolution detection boxes O that did not match successfully h-r Perform IOU matching and calculate the inactive tracking box K. uncomfirmed High-resolution detection boxes O that did not match successfully h-r The IOU cost matrix C3 is used to perform association matching between C3 and the adaptive matching threshold h3, resulting in the matched inactive tracking boxes K. uncomfirmed-m With high-resolution detection box O h-rm Unmatched inactive tracking box K uncomfirmed-r High-resolution detection boxes that did not match successfully h-rr The high-scoring detection bounding box O that was successfully matched h-rm Update the inactive tracking box K using Kalman filtering uncomfirmed-m The activated tracking box K is then obtained. uncomfirmed-m (Also called the posterior tracking box K) uncomfirmed-m ); VI. Anchor point rescue matching: High-resolution detection boxes O that failed to match in step five. h-rr Prior tracking box K that did not match successfully in step four rr Matching is performed. Since this is multi-class, multi-object tracking, the prerequisite for this step is matching within the same class to obtain the matched prior tracking boxes (trajectories) K. rrm With high-resolution detection box O h-rrmUnmatched prior tracking box K rrr High-resolution detection boxes O that did not match successfully h-rrr The high-scoring detection bounding box O after successful matching h-rrm The prior tracking box K is updated using Kalman filtering. rrm The new posterior tracking box K is then obtained. rrm ; The matching degree is determined as follows: The number of key points of the unmatched high-resolution detection boxes located within the range circle of the unmatched prior tracking boxes in step four is obtained. These key points are the four corner points and the center point of the high-resolution detection box. The range circle of the prior tracking box has the larger of its width and height as its radius and the center point of the prior tracking box as its center. This effectively increases the matching range of the prior tracking box. Figure 2 As shown in the image, the red box represents the tracking box, the blue box represents the detection box, the blue dashed line represents the occluded part of the detected object, the gray area represents the occlusion, and the green area represents the matching range of the tracking box. It can be seen from the image that before the improvement, the matching range of the tracking box was the area within the red rectangle; after the improvement, the matching range of the tracking box is the matching range of the green circle. The improved version increases the matching range compared to the original, and the circular matching range also better adapts to the problem of large target movements and frequent occlusion. Then, the number of keypoints within the tracking box's matching range that contain the detection box is multiplied by... The score is used as the matching criterion, with higher scores indicating higher matching accuracy. Here, d represents the distance between the center points of the high-scoring detection boxes that failed to match in step five and the prior tracking boxes that failed to match in step four, and the decay rate is controlled by the parameter 1 / 20. Traditional IOU cost matrices often lead to target loss due to low or even zero IOU values ​​between detection and tracking boxes when addressing the aforementioned problems. This method uses a score-based mechanism instead of the traditional IOU cost matrix as the matching criterion, allowing for more efficient utilization of high-scoring detection boxes. Specific matching rules effectively solve the problems of target occlusion and nonlinear motion. The original ByteTrack algorithm uses Intersection over Union (IOU) matching, which works by calculating the Intersection over Union (IoU) ratio of the detection box and the tracking box—the ratio of the area of ​​their intersection to their area of ​​their union. A higher IoU value indicates greater similarity between the two boxes. However, this method is often ineffective when considering boxes with angles. Therefore, the matching criterion in anchor point salvage matching is changed to checking how many of the four corner points and the center point of the high-scoring detection box are within the circle of the tracking box that failed the second matching attempt (using the larger of the tracking box's width and height as the radius, and the center point as the center). This effectively increases the matching range of the tracking box. Figure 2 As shown.

[0024] 7. Retention and Creation: Prior tracking boxes that failed to match in step 6 are retained by the ByteTrack algorithm for 30 frames and subjected to IOU matching upon their reappearance. Prior tracking boxes with more than 30 consecutive failed matches are discarded, while those with no more than 30 consecutive failed matches are directly used as posterior tracking boxes. High-resolution detection boxes that failed to match in step 6 are recreated as new tracking boxes K by the ByteTrack algorithm. n Newly created tracking boxes are directly used as posterior tracking boxes; retained prior tracking boxes are placed into the lost tracking box K. loss In the middle, the posterior tracking box K m Posterior tracking box K rm Activate tracking box K uncomfirmed-m Posterior tracking box K rrm Create a new tracking box K n These are placed in the posterior tracking bounding box K. tracked In the middle, the posterior tracking box K tracked The active tracking box is the one used to display on the current video frame.

[0025] 8. Repeat the above steps until there is no more video data.

[0026] The IOU matching method in steps three, four, and five above is to calculate the corresponding IOU. angle The cost matrix is ​​used, and the Hungarian algorithm is employed to perform association matching based on the cost matrix and an adaptive matching threshold. The IOU... angle The cost matrix is ​​as follows:

[0027] Among them, |θ detect -θ track | represents the difference in angle between the corresponding detection box and the tracking box; The formula for the adaptive matching threshold is as follows: .

[0028] To reflect the impact of target rotation angle on matching in the cost matrix, |θdetect-θtrack|, i.e., the detection box θ detect With tracking box θ trackThe angle difference. Since the above involves multi-target, multi-class detection and tracking, the cost value between different classes is directly set to 1. The angle difference between the detection and tracking boxes of different classes is no longer used in the cost matrix to prevent changes in the tracking box category. Because multiplying the cost matrix by α effectively reduces its size, the corresponding thresholds h1, h2, and h3 in the matching stage also change accordingly. Here, it is determined by taking the value from |θdetect-θtrack| that is actually used in the cost matrix and where the distance between the center points of the detection and tracking boxes is less than a certain distance (half the larger of the detection box's length and width plus half the larger of the tracking box's length and width). The larger of the two values ​​is used to modify the adaptive matching threshold. Through the above optimizations, the inherent limitations of traditional ByteTrack in rotating target scenarios are overcome, enabling the tracking system to not only maintain high accuracy under normal conditions, but also achieve more robust, stable and reliable rotating multi-target tracking in extremely complex environments such as underground coal mines.

[0029] The pseudocode for ByteTrack in this embodiment is as follows:

[0030] Comparative Example The steps of the existing ByteTrack algorithm are as follows: 1. Object detection: Target detection is performed on video data using the ByteTrack algorithm; 2. Classify the detection results: Based on the confidence score, classify the detection results by setting two thresholds, s1 and s2, and classify the detection result bounding box O as the high-scoring detection box O. h and low-resolution detection boxes O l Detection boxes with excessively low scores are discarded. High-scoring detection boxes correspond to more accurate detection targets, while low-scoring detection boxes contain false detections or targets that are not detected accurately. The posterior tracking bounding box K obtained from the previous frame of the video data (tracking trajectory) tracked Based on whether it is activated, it is divided into activation tracking box K. active Inactive tracking box K uncomfirmed Place the posterior tracking box that is in a lost state and has not been lost more than 30 times into the lost tracking box K. loss In the middle, activate the tracking box K active With the lost tracking frame K loss The prior tracking box K of the current frame is obtained through Kalman filter prediction. all ; III. Initial matching: high-resolution detection bounding boxes O h With prior tracking box K all The IOU is calculated to obtain the IOU cost matrix C1. The Hungarian algorithm is then used to perform association matching based on C1 and the matching threshold h1. After matching, the matched prior tracking boxes K are obtained. m With high-resolution detection box O h-m Unmatched prior tracking box K r High-resolution detection boxes O that did not match successfully h-r， The high-resolution detection bounding box that was successfully matched O h-m The prior tracking box K is updated using Kalman filtering. m The new posterior tracking box K is then obtained. m ; IV. Secondary matching: Prior tracking boxes K that did not match successfully in the initial matching. r With low-resolution detection box O l Perform IOU matching, calculate the IOU cost matrix C2 between the two, and use the Hungarian algorithm to perform association matching based on C2 and the adaptive matching threshold h2 to obtain the matched prior tracking boxes (trajectories) K. rm With low-resolution detection box O l-m Unmatched prior tracking box K rr Low-scoring detection boxes O that did not match successfully l-r The low-scoring detection box O after successful matching l-m The prior tracking box K is updated using Kalman filtering. rm The new posterior tracking box K is then obtained. rm ; 5. Three matches, tracking box K not activated. uncomfirmed High-resolution detection boxes O that did not match successfully h-r Perform IOU matching and calculate the inactive tracking box K. uncomfirmed High-resolution detection boxes O that did not match successfully h-r The IOU cost matrix C3 is used to perform association matching between C3 and the matching threshold h3, resulting in the matched inactive tracking boxes K. uncomfirmed-m With high-resolution detection box O h-rm Unmatched inactive tracking box K uncomfirmed-r High-resolution detection boxes that did not match successfully h-rr The high-scoring detection boxes that match successfully will be O h-rm Used to update inactive tracking boxes K uncomfirmed-m and the inactive tracking box K uncomfirmed-m The status is updated to active. VI. Retaining and creating prior tracking boxes K that failed to match above. rrThe ByteTrack algorithm retains 30 frames and performs IOU matching when the frame reappears. Prior tracking boxes with more than 30 consecutive failed matches are discarded, while prior tracking boxes with no more than 30 consecutive failed matches are directly used as posterior tracking boxes. The high-scoring detection boxes that failed to match are recreated as new tracking boxes K by the ByteTrack algorithm. n The newly created tracking bounding box is directly used as the posterior tracking bounding box; for the retained prior tracking bounding box K... rr Place it in the lost tracking box K loss In the middle, the posterior tracking box K m Posterior tracking box K rm Activate tracking box K uncomfirmed-m Posterior tracking box K rrm Create a new tracking box K n These are placed in the posterior tracking bounding box K. tracked In the middle, the posterior tracking box K tracked The active tracking box is the one used to display on the current video frame.

[0031] 7. Repeat the above steps until there is no more video data.

[0032] In this comparative example, the matching threshold is a fixed value set manually, and the cost matrix is ​​as follows: Cost = 1 - IOU As can be seen from the above embodiments and comparative examples, this invention addresses the problems of discontinuous detection results, transient missed detections, and even target loss in drilling pipe counting in coal mines based on rotating target detection, which affect the reliability of footage statistics. It overcomes the inherent limitations of ByteTrack in rotating target scenarios in existing technologies, enabling the tracking system to maintain high accuracy under normal conditions and achieve more robust, stable, and reliable multi-target tracking in the extremely complex environment of underground coal mines. By ensuring the target's identity remains unchanged and its trajectory is continuous over time, it significantly improves the stability and reliability of the data upon which footage statistics depend, providing stable detection results for subsequent drilling pipe counting. Specifically, this invention introduces the previously failed matching tracking box K after the ByteTrack processing has not determined the tracking box. rr High-scoring detection boxes O that failed to match h-rr Geometric anchor point salvage matching is performed using distance matching based on the corner and center points of the tracking box and the detection box. The angle difference between the detection box and the tracking box is introduced into the matching cost matrix, and an adaptive threshold based on angle difference perception is added to dynamically adjust the matching conditions, thereby improving the association stability and robustness in rotation and jitter scenarios.

[0033] To address the continuous sensing requirements of rotating targets in underground drilling rig operations, a robust multi-target tracking system based on YOLOv11-OBB and an improved version of ByteTrack is constructed. First, the next-generation backbone and decoupled head structure of YOLOv11-OBB are utilized to improve the angle estimation accuracy and geometric fitting capability of components such as drill pipes and drill tools. Second, considering the nonlinear motion and frequent occlusion characteristics of underground scenarios, the matching mechanism of ByteTrack is improved: on the one hand, new association indicators are constructed by introducing the length, width, center point, and geometric relationships of the rotating bounding box, improving the utilization rate of high-resolution detection boxes; on the other hand, a new cost matrix based on the rotating bounding box is designed, allowing the original cost matrix to consider the influence of the angle difference between the detection box and the tracking box, and an Angle-aware Adaptive Threshold (AAT) is added to dynamically adjust the threshold of the association stage based on the angle difference between the detection box and the tracking box. Overall, this method overcomes the limitations of traditional ByteTrack in rotating target scenarios, achieving more robust, continuous, and reliable multi-target tracking under complex underground conditions. Maintaining the consistency of target identity and trajectory continuity provides more reliable data support for borehole footage calculation.

[0034] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A continuous tracking method for coal mine drilling rods based on geometric correlation and angle threshold, characterized by constructing a tracking framework suitable for rotating targets based on the ByteTrack algorithm. Includes the following steps:

1. Object detection: Target detection is performed on video data using the ByteTrack algorithm; 2. Classify the detection results. Classify the detection results according to the confidence score. Set a threshold to divide the detection results into high-scoring detection boxes and low-scoring detection boxes. High-scoring detection boxes correspond to more accurate detection targets, while low-scoring detection boxes contain false detections or targets that are not detected accurately. The posterior tracking boxes obtained from the previous frame of video data are divided into active tracking boxes, inactive tracking boxes, and missing tracking boxes. The active tracking boxes and missing tracking boxes are processed to obtain the prior tracking boxes.

3. Initial matching: Perform IOU matching between high-scoring detection boxes and prior tracking boxes. Update the prior tracking boxes with the successfully matched high-scoring detection boxes using Kalman filtering to obtain the posterior tracking boxes. Fourth, secondary matching: the prior tracking boxes that failed to match in the first matching are matched with the low-scoring detection boxes using IOU matching. The successfully matched low-scoring detection boxes are then updated with Kalman filtering to obtain the posterior tracking boxes.

5. Three-way matching: Inactive tracking boxes are matched with high-scoring detection boxes that have not been matched successfully using IOU. The high-scoring detection boxes that have been matched successfully are then updated with Kalman filters to obtain posterior tracking boxes.

6. Anchor point rescue matching: The high-scoring detection boxes that failed to match in step 5 are matched with the prior tracking boxes that failed to match in step 4. The high-scoring detection boxes that successfully match are updated with Kalman filtering to obtain the posterior tracking boxes.

7. Retention and Creation: For prior tracking boxes that failed to match in step 6, the ByteTrack algorithm retains them for 30 frames and performs IOU matching when they reappear. Prior tracking boxes with more than 30 consecutive failed matches are discarded, while prior tracking boxes with no more than 30 consecutive failed matches are directly used as posterior tracking boxes. High-resolution detection boxes that failed to match in step 6 are newly created as inactive tracking boxes by the ByteTrack algorithm, and inactive tracking boxes are directly used as posterior tracking boxes.

8. Repeat the above steps until there is no more video data.

2. The method for continuous tracking of drill rods in coal mine drilling based on geometric correlation and angle threshold as described in claim 1, characterized in that, The matching degree determination in step six is ​​as follows: The number of key points of the unmatched high-scoring detection boxes located within the range circle of the unmatched prior tracking boxes in step four is calculated by multiplying the number of points by... The score is used as the matching basis, and the higher the score, the higher the matching degree; where d is the distance between the center point of the high-scoring detection box that failed to match in step five and the center point of the prior tracking box that failed to match in step four.

3. The method for continuous tracking of drill rods in coal mine drilling based on geometric correlation and angle threshold as described in claim 2, characterized in that, The key points are the four corner points and the center point of the high-resolution detection frame.

4. The method for continuous tracking of coal mine drilling rods based on geometric correlation and angle threshold as described in claim 2, characterized in that, The radius of the prior tracking box is the larger of its width and height, and the center of the prior tracking box is the center of the circle.

5. The method for continuous tracking of drill rods in coal mine drilling based on geometric correlation and angle threshold as described in claim 1, characterized in that, The IOU matching method in steps three, four, and five is to calculate the corresponding IOU. angle The cost matrix is ​​used, and the Hungarian algorithm is employed to perform association matching based on the cost matrix and an adaptive matching threshold. The IOU... angle The cost matrix (Cost) is as follows: Among them, |θ detect -θ track | represents the difference in angle between the corresponding detection box and the tracking box; The formula for the adaptive matching threshold is as follows: 。

Images