Mask Guided Pedestrian Tracking from Single Point Initialization

Abstract

Pedestrian tracking has a wide range of applications, such as intelligent monitoring and unmanned driving. In pedestrian tracking system, object detection algorithms are generally used to provide the initial bounding boxes for object tracking algorithms. If the bounding box provided by detection algorithm is not accurate, it will seriously affect the performance of object tracking. To solve this problem, this paper proposes a mask guided pedestrian tracking algorithm based on Siamese network. Firstly, Mask R-CNN is introduced to acquire exemplar image and its mask. Secondly, a light-weight convolutional neural network (CNN) is applied as the backbone of proposed algorithm. Thirdly, a channel attention module is introduced to our proposed algorithm and integrated with the light-weight CNN. The feature maps of the exemplar image and its mask are adjusted by a channel attention model and fused to enhance the discrimination ability of exemplar branch. Finally, our proposed tracker is trained GOT-10k dataset and evaluated on object tracking benchmark (OTB), experimental results show the excellent performance of proposed algorithm compared with state-of-the-art trackers.