Research On Object Tracking Algorithm Under Complex Background Using Anchor-free Siamese Network

Object tracking is an important research direction in the field of CV.Deep neural network is also used in more and more tracking algorithms.In this paper,a single object tracking algorithm based on anchor-free SNN with good resistance to complex background interference on a small low computing power embedded platform was studied.Aiming at the complex background problems that restrict the actual tracking effect,such as illumination change,occlusion and similar objects,the influence mechanism was analyzed and the benchmark algorithm Siam RPN++ was improved accordingly.Considering the balance requirements of network accuracy and complexity for model deployment under embedded platform,an optoelectronic tracking experimental platform based on embedded system was built to verify the tracking performance in the actual scene.The specific work of this paper is as follows:(1)A cross-scale feature extraction network and multi-scale cross-correlation method were studied.Aiming at the problems of weak discrimination of features and lack of semantic information under complex background,the asymmetric convolution operator including multi-scale convolution kernel was introduced to adapt to features with different aspect ratios,and the channel attention method was used to improve the cascade weight association of multi-layer features.The spatial dropout strategy was used to improve the effective utilization of fused features.UAVDT dataset with higher background interference frequency was used for testing.The success and precision rate were improved by 4.8% and 3.9% respectively.(2)An anchor-free output branch structure of tracking network was studied.Using the anchor-free idea,the regression and classification branches of base algorithm were redesigned.A global perception module based on spatial attention was introduced to realize the early separation of regression and classification features and independent background interference perception;Deformable convolution was used to explicitly align regression and classification feature vector,which improved the feature extraction ability of output branches in complex dynamic background.Compared with benchmark algorithm,the EAO of anchor-free branch structure under VOT2018 dataset was improved by 4.9%,and the post-processing time was reduced by 43.2%.(3)The photoelectric tracking experimental device based on small low computing power embedded platform was built.The overall improved algorithm and experimental platform were tested.Firstly,through qualitative and quantitative comparison with various tracking models,it was proved that the improved method could achieve optimal or suboptimal tracking effect on multiple datasets,meanwhile remain good tracking accuracy and reliability under complex background interference;Then the AI computing platform of the tracking experimental device is built based on the small low computing power embedded system under ARM+NPU architecture.After the hardware module assembly and algorithm transplantation and optimization,a scaled model was used to verify the actual performance under complex background interference.The experimental results showed that the photoelectric tracking experimental device can handle 1080p@30fps real-time video stream,which is suitable for object tracking applications under complex background interference.