High-speed moving target tracking of multi-camera system with overlapped field of view

Multi-camera target tracking is a research hot spot in the field of computer vision. Because the field of view covered by the fixed monocular camera is limited, it cannot continuously track fast-moving targets for a long time, such as the video sequence of skaters’ short-track speed skating. However, for such video sequences, the relative displacement of target between frames is relatively large and the advanced tracking algorithm is difficult to achieve accurate tracking, which has a considerable impact on the subsequent target handover. In view of this situation, we use four fixed cameras to build a tracking system to continuously track skater and propose to use convolutional neural network to extract the semantic features of the target to improve the accuracy of the target handover process. Aiming at this kind of fast motion video sequence with a fixed field of view, a new tracking algorithm was proposed. We establish the motion model of the target and integrate the speed information of the target into the sample generation mechanism. In addition, we also construct a scale filter in context to constantly and iteratively update the scale change information of the prediction box to adapt to the significant change of the visual properties of the target. It improves the stability and robustness of the system. Extensive experimental results in object tracking benchmark and our dataset of skaters illustrate outstanding performance of our method compared with the state-of-the-art methods, especially against the fast motion sequences with a fixed field of view.