Long-term tracking based on spatio-temporal context

Aiming at the problem that the fast tracking algorithm using spatio-temporal context (STC) will inevitably lead to drift and even lose the target in long-term tracking, a new algorithm based on spatio-temporal context that integrates long-term tracking with detecting is proposed in this paper. We track the target by the fast tracking algorithm, and the cascaded search strategy is introduced to the detecting part to relocate the target if the fast tracking fails. To a large extent, the proposed algorithm effectively improves the accuracy and stability of long-term tracking. Extensive experimental results on benchmark datasets show that the proposed algorithm can accurately track and relocate the target though the target is partially or completely occluded or reappears after being out of the scene.     

Branch-Activated Multi-Domain Convolutional Neural Network for Visual Tracking

Convolutional neural networks (CNNs) have been applied in state-of-the-art visual tracking tasks to represent the target. However, most existing algorithms treat visual tracking as an object-specific task. Therefore, the model needs to be retrained for different test video sequences. We propose a branch-activated multi-domain convolutional neural network (BAMDCNN). In contrast to most existing trackers based on CNNs which require frequent online training, BAMDCNN only needs offline training and online fine-tuning. Specifically, BAMDCNN exploits category-specific features that are more robust against variations. To allow for learning category-specific information, we introduce a group algorithm and a branch activation method. Experimental results on challenging benchmark show that the proposed algorithm outperforms other state-of-the-art methods. What’s more, compared with CNN based trackers, BAMDCNN increases tracking speed.