Online Multi-object Tracking with Multi-level Appearance Feature and Temporal Attention Mechanism

In this paper, a novel algorithm that integrates multi-level appearance feature and temporal attention mechanism is proposed for online multi-object tracking (MOT). To avoid identity switches and track fragment problems, low-level and high-level appearance feature extracted from different layers of a light-weighted convolutional neural network is employed to make the tracking more robust. Moreover, a new temporal attention mechanism is proposed to handle the problems of serious occluded and noisy detected objects. Specifically, temporal attention calculated from detection score, motion affinity and shape affinity are adaptively assigned to different samples in the tracklet. Third, a new feature updating strategy is applied to update the historical feature of samples in tracklet for long-term tracking instead of conventional approach such as directly removing the previous feature. From the experimental results, our method can achieve good tracking performance compared to start-of-art tracking algorithms, especially in identity consistency. Further, thanks to the design of light-weighted CNN and reused metric in temporal attention calculation, the proposed MOT can run up to SO Hz implemented on GTX1080.