Efficient Siamese model for visual object tracking with attention-based fusion modules

Abstract

In the realm of visual object tracking, achieving precise target localization poses a significant challenge. While Siamese trackers based on shallow networks often prioritize speed, they tend to suffer from subpar localization performance. This can be attributed to the inadequacy of features extracted by shallow networks, making it challenging to accurately pinpoint the target within intricate backgrounds. This paper presents an efficient Siamese model architecture for visual object tracking, aimed at mitigating the challenge of suboptimal feature extraction often observed in shallow network designs. On the basis of the Siamese network, we devise the Channel Attention Feature Enhancement Fusion Module (CAFEF) and Spatial-Channel Attention Feature Enhancement Fusion Module (SCAFEF) to bolster feature extraction while suppressing background noise. Furthermore, we enhance the model's resilience to disturbances by combining Siamese classification scores with those from an online classifier and to maintain high-quality image features during tracking, we introduce an online updating strategy for refreshing the template image. Experimental results across four publicly available datasets consistently demonstrate the outstanding performance of the proposed Siamese model architecture.