The Correlation Filter with Adaptive Spatial and Temporal Regularization for Inland Ship Tracking

Ship tracking is an important task of inland waterway video surveillance. Inland waterway scenes are complex. Generic algorithms applied directly to inland river scenes are susceptible to performance degradation due to boat occlusion, light changes, and water ripples. In this paper, we propose the ATSR-DCF (Self-adaptive Temporal and Spatial Regularization Discriminative Correlation Filter) algorithm to improve ship tracking performance using adaptive spatiotemporal regularization and ship position incremental information. First, ATSR-DCF gets the initial frame and initial spatial and temporal regularization weights to train the correlation filter. Second, input other video frames, compute the optimized spatial and temporal regularization weights and update the filter using the Alternating Direction Method of Multipliers (ADMM). Finally, obtaining video sequence position increments constrains the subsequent position for predicting the target ship. In order to evaluate the performance of ASTR-DCF, we perform experiments using our group’s inland waterway ship dataset. The results show that the ATSR-DCF tracking performance outperforms other comparative algorithms. The tracking success rate is 80.0% and the accuracy rate is 86.6%.