Region-of-interest adaptive digital zooming method

Abstract

The invention discloses a self-adaptive digital zooming method for a region of interest, and belongs to the field of computer vision and computer control. According to the invention, non-linear normalization is carried out on the target miss distance and the target aspect ratio, self-adaptive zooming is carried out on the region of interest containing the moving target, single-target detection and tracking in a high dynamic distance range under high resolution and large field of view are realized, and the space and time resolution of target tracking is improved. The method is mainly applied to the fields of intelligent transportation, intelligent shooting and the like, and can solve the problems that a single sensor is difficult to consider large view field and long distance and is low in tracking precision.

Description

Technical Field

[0001] The present invention relates to a method for adaptive digital zooming of regions of interest, belonging to the fields of computer vision and computer control. Background Art

[0002] In recent years, with the development of computer vision, the single-object detection and tracking (SODT) technology has important applications in the fields of video surveillance, target analysis, intelligent ice and snow sports, etc., and thus has received extensive attention from scholars and the industry at home and abroad.

[0003] A wide viewing angle and high resolution have always been irreconcilable problems in machine vision. Therefore, achieving a trade-off between spatial resolution and temporal resolution over a long distance and a large range is still a key challenge in machine vision. For large-scale scene applications, existing methods are mainly target detection and tracking methods based on regions of interest (ROIs). ROI detection can reduce noise interference in non-target regions. For example, in the field of medical imaging, remote photoplethysmography, fingerprint and hand vein detection, and blink detection for monitoring patients with amyotrophic lateral sclerosis, where only high-quality diagnosis-related regions are required. For computationally intensive fields, transmitting ROI images in aerial surveillance can significantly reduce the data transmitted by aerial drones, and reconstructing only on the ROI will only significantly speed up the processing. Similarly, in lane detection applications, processing only a smaller ROI instead of the entire image can achieve faster detection, which is crucial for a real-time vehicle warning system. The target tracking method based on correlation filtering also applies the ROI technology. The object of interest in the video frame is detected by a neural network, and then the position of the ROI in the next frame is predicted using an estimation filter. However, fixed ROIs often reduce the scope of these methods. Therefore, the selection strategy of ROIs is also a key challenge in large field-of-view and high-resolution scenarios. Summary of the Invention

[0004] To solve the above problems, the object of the present invention is to propose a method for adaptive zooming of regions of interest for tracking and photographing high-speed moving targets at a long distance. This method performs adaptive ROI digital zooming through a large field-of-view vision sensor, then applies an advanced target detection and tracking algorithm to detect and track the target, and at the same time updates the RoI in real time based on the detected results, and sends the tracking result as a control signal to the P / T turntable through a control algorithm to drive the turntable to complete real-time tracking of the target.

[0005] The object of the present invention is achieved by the following technical solutions.

[0006] A method for adaptive zooming of regions of interest for tracking and photographing high-speed moving targets at a long distance disclosed by the present invention includes the following steps:

[0007] Step 1: Divide the image of the entire FOV captured by the vision sensor into m×n partitioned images, and perform cyclic detection on the partitions through a detection algorithm;

[0008] Step 2: If an object is detected in Step 1, the algorithm will generate an ROI region based on the detection result and co-predict the ROI of the next frame according to the detection and tracking results. Among them, ROI ≤ FOV, t loop is the time for one traversal detection of the partitions, t detect is the time required to detect the target, with the unit of s / frame, t track is the time required by the actuator;

[0009] Step 3: During the detection process in Step 1, if situations such as target loss, target interference, and target change occur, the adaptive ROI update algorithm will adaptively update the ROI region through digital zoom according to the target tracking situation. The formula is as follows,

[0010] ROI new = S(ROI old , w, h)+βD(ROI old , x, y)

[0011] Among them, ROI old is the previous ROI region, w and h are the width and height of the target box respectively, and x and y are the center positions of the target box. The ROI adaptive digital zoom algorithm based on the normalization of the center point distance and aspect ratio change is calculated as follows,

[0012]

[0013] Among them, b = [x, y] and b i-1 = [x i-1 , y i-1 are the center points of the target boxes in adjacent images, c is the diagonal length of the target box, ρ is the Euclidean distance, and α is the size offset factor.

[0014] Step 4, during the target tracking process, set the target detection failure delay threshold t delay . When the target loss time is greater than t delay , then the target is lost for a long time, and the position change factor γ is calculated by the improved Gaussian distribution as follows,

[0015]

[0016] Among them, b cen = [x cen , y cen is the center point of the image, σ is the target deviation variance.

[0017] The adaptive zoom method for the region of interest (ROI) in case of target loss is calculated as follows:

[0018]

[0019] Step 5: Perform stable adaptive update of the ROI position by non-linearly normalizing the center point distance, so that the position of the ROI can also adaptively change with the detection result, improving the stability of the ROI position change.

[0020]

[0021] Step 6: Match the target to be tracked. Obtain a new target to be tracked through the target detection algorithm, and calculate new target tracking information through the target tracking algorithm. Then continue to execute the adaptive zoom method for the region of interest described in Steps 2 to 6.

[0022] Beneficial effects:

[0023] The adaptive zoom method for the region of interest for tracking cameras of high-speed moving targets at a long distance disclosed by the present invention adopts an ROI adaptive update model based on the normalization of the center point distance and aspect ratio change, and is used for single target detection and tracking in a high-dynamic distance range under high resolution and a large field of view. It can maintain a high processing efficiency of time resolution within a large field of view of high resolution, and effectively alleviate the problem that it is difficult for a single sensor to balance the large field of view and long distance and has low tracking accuracy when tracking high-speed moving targets at a long distance. Brief description of the drawings

[0024] Figure 1 is the algorithm flowchart of the tracking method for ROI adaptive digital zoom disclosed by the present invention;

[0025] Figure 2 is the schematic diagram of the ROI adaptive digital zoom system of the present invention; Detailed implementation manners

[0026] To clearly illustrate the technical solution proposed by the present invention, the present invention will be described in detail below with reference to the drawings and specific embodiments. It should be noted that the described examples are only intended to facilitate the understanding of the present invention and do not limit it in any way.

[0027] As Figure 1 shown, for the tracking method of ROI adaptive digital zoom in this embodiment, the specific implementation steps are as follows:

[0028] Step 1: Divide the image of the entire FOV captured by the visual sensor into m×n partitioned images, and perform cyclic detection on the partitions through the detection algorithm;

[0029] Specifically, using a large field of view camera as the vision sensor, the entire field of view captured by the vision sensor is divided into m×n partitioned images. Subsequently, the advanced Yolov8 object detection algorithm is applied to detect region by region. When no object is detected, the detection continues to loop from the first partition until the target to be tracked is detected. The variable particle filter algorithm based on image measure metric resampling is used to process the target miss distance, and the control signal is sent to the P / T turntable through the control algorithm to drive the turntable to complete the real-time tracking of the target.

[0030] In addition, in this embodiment, to ensure the detection and tracking of large sites within a distance of 0 - 170m, we selected a large field of view industrial camera (MS-XG903C / M) with high resolution and high frame rate as the vision sensor according to the single aperture imaging model. Its resolution is 4208×2160, the field of view angle is 40.2×30.6, and the imaging frame rate is 100fps.

[0031] Step 2, if an object is detected in Step 1, the algorithm will generate an ROI region based on the detection result and jointly predict the ROI of the next frame according to the detection and tracking results. Among them, ROI ≤ FOV, t loop is the time for one partition traversal detection, t detect is the time required to detect the target, with the unit of s / frame, t track is the time required for the actuator.

[0032] Step 3, during the target tracking process, if there are situations such as background interference and target changes, the adaptive ROI update algorithm will adaptively update the ROI region by digital zoom according to the target tracking situation.

[0033] ROI new = S(ROI old , w, h)+βD(ROI old , x, y)

[0034] Among them, ROI old is the previous ROI region, w and h are the width and height of the target box respectively, and x and y are the center positions of the target box respectively. The normalized center point distance and aspect ratio change are used in the ROI adaptive digital zoom algorithm to improve the stability of the algorithm.

[0035]

[0036] Among them, b = [x, y] and b i-1 = [x i-1 , y i-1 are the center points of the target boxes in adjacent images, c is the diagonal length of the target box, ρ is the Euclidean distance, and α is the size offset factor.

[0037] Step 4, during the target detection process, if the target is lost for a long time, set the target detection failure delay threshold t delay . When the target loss time is greater than t delay , to prevent the target from being lost in the ROI, the ROI area should increase as the target loss time increases. Since in the tracking state, the ROI area always moves with the target, and the target is in the center area of the detection screen, its position distribution in the screen is similar to a Gaussian distribution. To prevent the ROI area from being too large to detect the target, an ROI size adaptive update method based on time change and position change is proposed, and the position change factor γ is calculated by the improved Gaussian distribution. The greater the deviation of the target from the center position in the previous frame, the greater the reduction of the ROI, and conversely, the smaller the deviation of the target from the center position in the previous frame, indicating that the target is stable, and the reduction of the ROI is appropriately smaller.

[0038]

[0039] where b cen = [x cen , y cen is the center point of the screen, σ is the target deviation variance.

[0040] Therefore, the formula for adaptive ROI digital zoom is as follows:

[0041]

[0042] Step 5, perform stable adaptive update of the ROI position through non-linear normalized center point distance so that the position of the ROI can also adaptively change with the detection result, improving the stability of the ROI position change,

[0043]

[0044] Step 6, for the matching of the target to be tracked, obtain a new target to be tracked through the target detection algorithm, and calculate new target tracking information through the target tracking algorithm, then continue to execute the ROI adaptive zoom method described in Steps 2 to 6.

[0045] This embodiment tests the U-shaped field trick movement by combining an intelligent tracking shooting system, and can achieve stable tracking shooting of high-speed skiers.

[0046] In summary, the above are only the preferred embodiments of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for adaptive zooming of a region of interest for tracking and photographing a high-speed moving target at a long distance, characterized in that: The method comprises the following steps: Step 1: divide the image of the entire FOV captured by the visual sensor into m×n partitions, and use the target detection algorithm to perform cyclic detection on the partitions; Step 2: If an object is detected in step 1, the algorithm will generate a region of interest (ROI) based on the detection results and collaboratively predict the ROI of the next frame based on the detection and tracking results. loop is the time to perform a partition traversal detection, t detect is the time required to detect the target, in seconds per frame, t track It is the time required by the executive body; Step 3: During the target tracking process, if background interference and target changes occur, the adaptive ROI update algorithm will perform adaptive digital zoom updates on the ROI area according to the target tracking situation. The formula is as follows: KING new =S(KING) old ,w,h)+βD(ROI old ,x,y) Among them, ROI old is the previous ROI area, w, h are the width and height of the target frame, x, y are the center position of the target frame, and the ROI adaptive digital zoom algorithm based on the normalization of the center point distance and aspect ratio change is calculated as follows: Where b = [x, y] and b i-1 =[x i-1 ,y i-1 ] is the center point of the target box of the adjacent picture, c is the diagonal length of the target box, ρ is the Euclidean distance, and α is the size bias factor. Step 4: During the target tracking process, set the target detection failure delay threshold t delay , when the target loss time is greater than t delay When , the target is lost for a long time, and the position change factor γ is calculated by the improved Gaussian distribution as follows: Among them, b cen =[x cen ,y cen ] is the center point of the picture, and σ is the target deviation variance. The adaptive zoom method for the region of interest with target loss is calculated as follows: Step 5: Stably and adaptively update the ROI position by nonlinearly normalizing the center point distance so that the ROI position can also change adaptively with the detection results, thereby improving the stability of the ROI position change. Step 6: The target to be tracked is matched, a new target to be tracked is obtained through a target detection algorithm, new target tracking information is calculated through a target tracking algorithm, and the region of interest adaptive zoom method described in steps 2 to 6 is continued.

Images