Research On Siamese Network Tracking Algorithm Based On Feature Fusion

Visual object tracking is one of the fundamental problems in computer vision research.Given the bounding box of the target object in the first frame of the video,the tracker expects to locate the target object in all subsequent frames.The biggest challenge of visual tracking can be attributed to the requirement of robustness and recognition ability at the same time.Robustness is that the tracker does not lose the target when its appearance changes due to lighting,motion,perspective,or object deformation.At the same time,the tracker is required to have the ability to distinguish the target object from the cluttered background or similar surrounding objects.Although many breakthrough achievements have been made in the research on robustness and recognition ability in the tracking field,there are still barriers to the tracking effect in some complex scenes,especially in the tracking of fast moving objects.Therefore,the outstanding problems and challenges in the current target tracking field are studied,and the specific research results are as follows.This paper proposes a feature fusion-based Siamese network tracking algorithm,namely Optimized Spatial Matching for Visual Object Tracking(OSM-Tracker).The algorithm is a two-stage anchorless Siamese network tracker composed of an optimized spatial network and a correction network.This thesis mainly makes innovatioans in the following three aspects: 1.The optimized spatial fusion module is introduced in the feature extraction stage to learn how to filter redundant information in space to avoid inconsistencies and improve the scale invariance of features.2.The anchor-free tracking strategy is adopted in the optimized spatial network to avoid introducing too many anchor hyperparameters and anchor boxes,thereby reducing the cost of model training and operation.3.The convolution self-calibration module is introduced in the target correction stage.Through the innovative self-calibration convolution,the correlation between space and channels is established around each pixel point,which improves the information content of the output features and improves its representation learning ability.Through the cooperation of many aspects,OSM-Tracker has produced good results.Experiments show that the tracker designed in this paper achieves a large performance improvement on the three benchmarks of VOT-2017,OTB-100 and La SOT,and achieves real-time results.