In recent years, correlation filtering and deep learning have achieved good performance in object tracking. Correlation filtering is an efficient and real-time method because its formula provides a fast solution in the Fourier domain, but it does not benefit from end-to-end training. Although deep learning is an effective method for learning object representations, training deep networks online with one or a few examples is challenging. To address these problems, we propose a deep ensemble object tracking algorithm that fuses temporal and spatial information to improve algorithm precision and robustness. The framework of our algorithm includes four aspects: feature extraction, a baseline network, a branch network and adaptive ensemble learning. Feature extraction extracts the general object representation. The baseline network integrates feature extraction and a correlation filtering algorithm into a convolutional neural network for end-to-end training. The branch network is composed of a temporal network and a spatial network. The temporal and spatial networks capture the object temporal and spatial information and further refine the object position. Our algorithm only needs an initial frame to train all networks. Adaptive ensemble learning compensates for the object information deficiency and improves tracking accuracy. Many experiments on tracking benchmark datasets demonstrate that our algorithm performs favourably compared with state-of-the-art tracking algorithms.
Siamese trackers draw continuous attention in the object tracking community due to their proper balance between performance and inference speed. Despite that, it remains unclear how to effectively exploit the target appearance cues and motion cues involved in videos to improve trackers’ performance. To address this problem, we develop a Siamese network with diverse prior information integrated, namely DPINet, by extending two novel blocks to a powerful anchor-free Siamese network. First, we design a channel- and space-aware feature enhancement (CSE) block to highlight target-specific feature weights in two aspects (channel and spatial dimensions). It is devoted to making full use of the target cues in the initial frame by considering them as guidances, in which way target-related representation in feature maps can be improved. It also facilitates the interplay between two input branches. Second, we advance a cross-correlation block with multi-dimensional information fusion (MDI-XCorr). In this block, target motion cues within adjacent frames can be mined and treated as supervisions to refine the response map in the current frame during inference. Hence, both tracking quality and stabilization can be enhanced. Evaluations on four popular benchmarks are conducted, showing that DPINet achieves 0.702 (AUC), 0.474 (EAO), 0.336 (EAO), 0.613 (AO), and 0.527 (AUC) on OTB100, VOT2018, VOT2019, GOT-10k, and LaSOT, respectively.
6 May 2020 in IET Cyber-Physical Systems: Theory & Applications
Cyber physical system (CPS) is a complex system combining computation, network and physics; object tracking is an important application of CPS. To solve the problem that the traditional kernel correlation filtering tracking algorithm cannot recover the lost object, the authors propose a re‐detection object tracking algorithm. The proposed algorithm mainly designs a new adaptive detection criterion. By comparing the value of detection criterion and the value of the experience threshold, it can be judged whether the current target is lost or not. When the object tracking fails, the proposed method can generate target candidate boxes by using the edge boxes algorithm and select the best target location by applying the non‐maximum suppression and the Euclidean metric methods. In addition, a fast multi‐scale estimation method and an adaptive updating method are added to the tracking procedure to further improve the overall performance of the algorithm. Experimental results show that the proposed approach has a good performance in terms of precision and success rates.