Research On Online Object Tracking Method Based On Visual Long Short-Term Memory

With the development of computer hardware and machine learning theory,visual object tracking techniques have important applications in areas such as unmanned vehicles,video reconnaissance,and medical image analysis.However,the challenges such as frequent disappearance,target reappearance and appearance changes caused by occlusion in long-term tracking tasks can easily lead to tracking failure.How to adapt to these challenges is the key factor to achieve robust tracking.It is difficult for the template update strategy of traditional tracking methods by building a target appearance model to maintain accurate tracking results in long-term tracking scenarios.Therefore,the paper aims to achieve robust visual tracking in long-term complex scenarios by designing a memory network mechanism combined with deep learning techniques.The main research contents and innovation points of the paper are as follows.(1)Residual memory inference network for regression tracking with weighted gradient harmonized loss is proposed.Firstly,the residual memory inference network incorporating long short-term memory mechanisms is proposed for mining target history states.The single layer convolution is used as short-term memory for adapting to drastic appearance changes of the target.The Conv LSTM is combined in the form of residual learning to learn the attribute information of the target in the long-time state from a complementary perspective.The high-quality long short-term memories can express the state information under complex motion scenes of the target through the end-to-end learning.Secondly,to address the degradation of model performance due to the extreme imbalance of positive and negative samples,the method proposes a weighted gradient harmonized adaptive learning strategy.By introducing a gradient harmonized factor into the loss function of deep regression,the gradient criterion is used to balance the proportion of different samples in the loss to reduce the impact of data imbalance on model update learning,and the sample quality is evaluated to ensure that the memory network can learn reliable target information.Finally,the residual memory inference network and the weighted gradient harmonized adaptive learning strategy are unified into the target tracking framework to achieve accurate and robust target tracking performance.In the paper,the performance of the proposed algorithms is evaluated on public datasets such as OTB(Object Tracking Benchmark),TC-128(Temple Color),UAV-123(Unmanned Aerial Vehicles),and VOT(Visual Object Tracking).The experimental results demonstrate that the proposed tracker can effectively remember and utilize the target history state information,and is highly adaptive under the challenges of target occlusion and deformation.(2)Attention-driven memory network for online visual tracking is proposed.Firstly,to enhance the effectiveness of memory content,the attention-driven memory network is designed,and the long-term memory module obtains memory information at the level of essential attributes of the target by mining the state of the target in both channel and spatial dimensions.Meanwhile,the short-term memory module is added to the reciprocity to maintain desirable adaptability in the face of sharp deformation of the target.Under the weighted gradient harmonized loss,the attention-driven memory network can adaptively adjust the contribution of short-term and long-term memories to the tracking results.Secondly,the online memory updater is proposed to avoid the introduction of error information into memory leading to the degradation of model discrimination.By mining the target information in the tracking results online,the confidence level of the current results can be obtained,so that the timing of updating the memory can be accurately judged to ensure the effectiveness of the memory update.Finally,the attention-driven memory network is combined with the memory update discriminator to achieve robust and real-time visual tracking.The proposed method performs favorably and has been extensively validated on several benchmark datasets,including OTB-50/100,TC-128,UAV-123,GOT10 k,and VOT against several advanced methods.