Objectives Target tracking has been widely used in pedestrian tracking, automatic driving, target monitoring and other fields. The existing tracking methods in single camera have limited tracking range and relatively slow. The existing person re-identification methods rely on the pedestrian detection to establish image database, and all the methods may miss detection or lead to false detection. Therefore, in order to meet the needs of pedestrian tracking in large-scale scenes, a continuous pedestrian tracking method for vertical mounted camera is studied, which has better tracking robustness and faster speed.
Methods For tracking in a single camera, a discriminative correlation filter with detection and spatial reliability (DSR-DCF) was proposed. Firstly, the Gaussian mixture model was used to eliminate the background, and the minimum circumscribed rectangle of the target was extracted to select the tracking target. Then, 32 dimensional histogram of oriented gradient (HOG) feature and 1 dimensional grayscale feature are used as pedestrian feature description, and spatial reliability and detection reliability are applied to correlation filter to realize pedestrian tracking in single camera. In the process of tracking pedestrians across cameras, according to the imaging characteristics of pedestrians in the vertical mounted camera, the speeded up robust features (SURF) algorithm was used to match the features of overlapping areas. The homography matrix between adjacent shots was calculated according to the matching feature points to determine the best search area of tracking target in the adjacent camera. Finally, taking the pedestrian obtained by Gaussian mixture model background elimination in the search area and template pedestrian as input, and using the sum of absolute difference(SAD) as the matching measure, the real-time cross-lens continuous tracking of pedestrian was realized through template matching.
Results Scene simulation and tracking experiments were carried out with four cameras with a resolution of 1 920×1 080 pixels. The success rate was tested by online object tracking benchmark (OTB), and compared with kernelized correlation filters (KCF), discriminative correlation gilter with channel and spatial reliability (CSR-DCF) and other methods. The results show that the background elimination of Gaussian mixture model can extract all pedestrians, and there is no missing or false detection. When tracking across cameras, the best search range is 5 times of the initial tracking window size. In continuous tracking, the average speed of the proposed method can reach 21.8 frames/ s, and the tracking success rate is better, especially in the case of illumination change, deformation, complex background and occlusion.
Conclusions The single camera tracking method DSR-DCF, combined with search area restriction and template matching, can realize the continuous pedestrian tracking across camera. The tracking speed and success rate can meet the real-time requirements, and the tracking speed is better than 21 frames/s.
DownLoad: