Research On Correlation Filter Object Tracking With Visual Attention Mechanism

The spatially regularized correlation filter tracking algorithm has poor accuracy in the situations of target deformation,rotation and fast motion,and is also prone to tracking failure in abnormal situations such as occlusion.To address these issues,a correlated filtered object tracking algorithm with a visual attention mechanism is proposed with innovations around the way the model is constructed and the tracking anomaly handling strategy based on the baseline algorithm.First,the spatial attention mechanism is integrated into the algorithm.The saliency map of the target region is obtained using the saliency detection algorithm,and the spatial mask matrix is obtained after cropping and binarization of the saliency map,and then the spatial attention reference weights are fused with predefined spatial weights to generate spatial attention reference weights.Using this approach to construct spatial attention regular terms improves the algorithm’s adaptability in situations such as fast target motion.Next,the channel attention mechanism is integrated.Channel weights are introduced in the model and channel attention regularization term is constructed.The channel weights are optimized simultaneously in the training phase of the model to reduce the impact of redundant information in the multi-channel features on the tracking performance when the object is deformed and rotated.Finally,strategies for tracking anomaly handling are incorporated into the algorithm.In the model training stage,the change magnitude of the tracking response graph is constrained to decrease the interference when the target is blurred and other mild anomalies.In the model update stage,the Peak-versus-Noise Smoothness of the response map is analyzed to determine the occlusion of the tracking and filter low quality samples.It improves the stability of the model when the target is occluded and other serious anomalies.Four public datasets,OTB2015,TC-128,UAV123 and La SOT,are used in the experiments to compare with typical target tracking algorithms in recent years.Compared with the baseline algorithm,the proposed algorithm improves the accuracy by6.7%,6.0%,9.1% and 5.2%,and the success rate by 4.6%,5.3%,6.2% and 1.6%,respectively.The experimental results show that the proposed algorithm can effectively mitigate the interference of abnormal situations such as target occlusion to the tracker,and performs more robustly in complex scenarios such as target deformation and rotation.There are 29 figures,5 tables and 94 references in this paper.