Boosting the Speed of Real-Time Multi-Object Trackers
Xudong Zhang, Liang Zhao, Feng Gu · 2021
Object tracking is an important computer vision task that has drawn increasing attention because of its academic and commercial potential. Although several trackers based on convolutional neural network (CNN) have shown strong performance in single object tracking and multiple object tracking, it remains challenging to perform real-time object tracking on videos with frame rate over 30 frame per second. For real-world applications, one of the necessary requirements is the trackers need to estimate the bounding boxes with a rate higher or equal to the video frame rate to avoid latency. In this study, we propose a hybrid tracker combining the high accuracy of CNN-based trackers and the superior processing speed of particle filtering algorithm to achieve real-time tracking. The experimental results indicate our tracker achieves a state-of-the-art performance on the benchmark multi-object tracking datasets with competitive accuracy.