Multi-feature fusion-based robust tracking of small targets in unmanned surface vehicle vision

Abstract：Objectives To overcome the challenges of tracking small targets in unmanned surface vehicles vision under the conditions of low feature resolution and similar environmental information, a multi-feature fusion continuous convolution operator tracking algorithm is proposed. Methods The resolution of multi-feature maps is enhanced using bicubic interpolation techniques, enabling sub-pixel level localization. Efficiencies in target tracking are achieved through feature projection and sample space generation, mitigating filter overfitting. Furthermore, interference arising from similar environmental features on the filter is addressed by developing an update strategy for high-confidence models. Results Validation and comparative analysis were performed using several representative video datasets. Experimental results showed that, compared to traditional continuous convolution operators tracking algorithms, an average success rate increase of 17.4%, an average distance precision increase of 17.8%, and an expected average overlap rate increase of 5.1% were achieved. Conclusions The proposed algorithm can deal with the small target tracking confusion problem in marine environment, and provide key technical support to improve the intelligent sensing capability of unmanned boats and marine robots.