Abstract:

To address the performance degradation of target tracking algorithms caused by target object deformation, scale variation, fast motion, and occlusion, a highly robust target tracking algorithm that Merging a CNN and Transformer is proposed based on siamese network architecture. In the feature extraction stage, standard convolutions are employed to extract shallow local feature information, while a convolution-like Transformer module is designed in the deep network to model global information. The pixel values in the Transformer are computed using a sliding window significantly reducing computational complexity. In the feature aggregation stage, a multi-head cross-attention module is utilized to construct a network for feature enhancement and aggregation. This module filters outirrelevant information and highlights the template-related information to improve the discriminative power of the features. Compared with the current mainstream algorithms, the proposed algorithm is optimal in terms of evaluation metrics under four different challenges of deformation, scale variation, fast motion and occlusion, on OTB2015 dataset. The average overlap(AO) on GOT-10K dataset is 70.8%, which is an improvement of 3.7% and 5.9% compared to the TransT and SiamR-CNN algorithms, respectively. The success rate on LaSOT and UAV123 dataset is 67.7% and 71.9%, which is improved by 2.8%, 2.8% and 2.9% and 7% compared to TransT and SiamR-CNN algorithms, respectively. The robustness(R) evaluation results on the VOT2018 and VOT2019 datasets show that the proposed algorithm achieved the least tracking failures rate, with robustness(R) index scores of 0.112 and 0.266, respectively, which are improved by 0.5% and 5% compared to Ocean algorithm, further verifies the higher robustness of the proposed algorithm.

Key words: object tracking, siamese network, Transformer, multi-head cross-attention mechanism, high robustness