Research On Abrupt Motion Target Tracking Algorithm Based On Siamese Network

In recent years,the research of visual tracking algorithms is no longer confined to the use of traditional correlation filtering algorithms to extract target features,but through deep learning-based neural networks to extract the depth features with powerful expression capabilities.This breakthrough has provided strong support for the rapid development of the target tracking field,making some tracking algorithms achieve good performance,especially the tracking methods based on the Siamese network,derived from a large number of both speed and accuracy of tracking methods,and therefore become a popular research branch in the field of target tracking.However,tracking algorithms based on the Siamese network often lead to tracking drift or even failure under the challenges caused by abrupt motion such as fast motion,deformation,and occlusion,mainly because the trackers cannot adapt to the feature changes of the targets in time.The main research of this paper is to make the tracker adapt to the target feature change timely and switch the tracking process according to the target state,to reduce the failure probability of the tracker when facing the target abrupt motion.The main research and innovations of this paper are as follows:(1)An online bionic visual Siamese tracking based on a mixed time-event triggering mechanism is proposed.Due to the cumulative error in tracking results and the lack of a suitable model update strategy,the online update based Siamese tracking algorithm still suffers from model drift.To solve this problem,the bionic vision network introduces the receptive field block and the blurpool,which improve the quality of feature extraction while maintaining the translational invariance of the convolutional neural network.The former uses dilated convolution kernels with different dilation rates to fuse depth features,which effectively increases the receptive field of the network.The latter uses low-pass filtering to anti-alias before down sampling,reducing the negative impact of the down sampling operation on the generalization ability of the network.In addition,to enable the model to effectively capture target appearance variations,a template update strategy with the mixed time-event triggering mechanism is designed.The strategy evaluates the quality of tracking results via a quality assessment model,guided by the mixed time-event triggering mechanism to an adaptively weighted fusion of fixed and mutative templates.These enable the tracker to adapt to feature changes timely,maintaining stable tracking under difficult scenarios caused by the abrupt motion of the target.(2)A discriminant enhanced memory and reciprocal target-distractor proposals for abrupt motion tracking is proposed.Firstly,the discriminant enhanced memory model can better identify the target from potential proposals,which is updated online by multi-peak sample evaluation and Matthew effect loss.The former mines hard negative samples,and the latter encourages the model to focus on hard negative samples.Secondly,the question-guided interval spatio-temporal constraint strategy is designed,which adaptively adjusts the constraint weights on the response map according to the tracking state.In addition,the reciprocal target-distractor proposal method is designed,which filters a large number of indistinguishable proposals by common sense information and a reciprocal target-distractor constraint strategy.These enable the algorithm to adjust the tracking strategy timely according to the state of the target,adapt to the difficult scenarios caused by the abrupt motion of the target,and get a large improvement in performance.To evaluate the actual performance of the algorithm,several representative tracking datasets OTB100,UAV123,La SOT,VOT2016,VOT2018,and GOT-10 k are selected for performance testing in the current research area,and the experimental results show that the two tracking methods achieve better results in abrupt motion target scenarios,and the algorithm tracking performance is improved in several other types of challenging scenarios.