| High-speed videogrammetry is a new discipline that integrates the advantages of close-up photogrammetry and computer vision technology to measure the target object in real time by non-contact means.The sequence images captured by high-speed cameras can record the position and posture of the object at a moment of motion,and can also perform quantitative analysis on the object through photogrammetric analysis methods.The recorded images can also be used as measurement data for future research reuse.Circular mark act as an intermediary in high-speed videogrammetry,posting motion information at key locations of the object under test.However,due to the complex shooting scenes,small size of circular marks and many interfering information,it is difficult for existing traditional algorithms to accurately identify and track circular marks.In this paper,we study how to improve the recognition and tracking effect of circular marks,and the main research contents are.High accuracy circular mark recognition is of great significance for camera calibration,target tracking and 3D reconstruction.However,most of the existing studies mark a single simple scene.These aforementioned algorithms lead to higher false alarm and missed alarm rates in more complex contexts.In this paper,we present a highprecision recognition method based on a novel deep learning model,Circular-Mark Net(CMNet)to solve this problem.The proposed network consists of three parts: first,a modified YOLOv4 network is used to detect circular marks so as to narrow the search area of circular contours;then a saliency object detection model BASNet is used to extract the contours of circular marks;and finally,least square fitting(LSF)is used to calculate the central pixel coordinate of the identified contour on the saliency map.The proposed method was tested under three complex scenarios with different characteristics and disturbances.The experimental results demonstrated that:(1)the proposed CMNet can effectively recognize of circular marks within complex scenes,which reveals the superiority and generalization ability of the proposed method;(2)the improved YOLOv4 can significantly enhance the detection accuracy of circular marks,which is crucial to the subsequent saliency courter detection and circle center identification;(3)Circular-MarkNet had a pixel RMSE of 0.0713,surpassing the state-of-the-art methods,and becoming the best algorithm.Circular mark tracking is of great importance for high-speed videogrammetry.The conventional methods using grayscale based sub-pixel level matching method and phase correlation based sub-pixel level matching method need to manually frame the first frame image circle center position,which reduces the whole algorithm automation process.And it is difficult to perform effective tracking when encountering experiments with large target displacement.In this paper,we use the improved DeepSORT algorithm for circular mark tracking,using CMNet to replace the original detector,making the algorithm more focused on the detection of circular marks.Then,using the cascade strategy,the depth features of the detection frame IoU and the circular marks in the detection frame are used to construct the cost matrix for the best pairing of the detected trajectory and the predicted trajectory.The availability of the method applied to high-speed videogrammetry is demonstrated by validation on sequential images of circular mark. |
