Research On Object Tracking's Technology Based On Spiking Neural Network

As the "third-generation neural network",spiking neural network has biological plausible,and has the advantages of high computational efficiency,low energy consumption,low resource consumption,and easy hardware implementation.However,the bottleneck of the immature and inefficiency training algorithm,which limits the spiking neural network's application field.On the other hand,tracking technology is an important research direction in computer vision,and has widely used in different applications for security,automatic vehicles.But the short-term tracking model based on deep learning is difficult to apply to edge devices such as resource-constrained artificial intelligence chips due to the large amount of computation,the large amount of resources,and the need of more powerful computing platforms(e.g.GPU).Based on the limitations of spiking neural network and the application bottleneck of tracking technology,this thesis attempts to apply spiking neural networks to short-term tracking.The main research contents include:1.The RFT coding based on the receptive field and the AAR coding based on the attention mechanism are proposed for coding image into spike trans.And choose the coding method for specific application.2.A more comprehensive parameter migration scheme is proposed.With the processes of standardization,retraining,parameter migration,and weight normalization,we can transfer a trained convolutional neural network to a similar structure's spiking neural network.The transferred spiking neural network can achieve the accuracy of 0.976 on the handwritten digit recognition with the AAR coding and the optimal parameter setting.Compared with the original convolution neural network,the accuracy decreases by only 0.007.However,there are significant improvements in the performance of resources,calculation speed,power consumption.3.The WISI distance is proposed to evaluate the similarity of spiking feature.Compared with the ISI distance evaluation,The effect of the WISI assessment is more accurate with considering the precise firing times;4.Based on the full-convolutional siamese network and the research on coding?model construction and similarity,a short-range tracking model with spiking neural network named Siam SNN is proposed.In the early research,this model can achieves a certain success rate,and has the advantages of low energy consumption,and low resource consumption.These advantages ensure that it is more easier to apply Siam SNN to edge devices.The technology about applying spiking neural network to short-term tracking proposed in this theies,which can provide reference value for spiking neural network's future application in other computer vision areas,and have strong practical value.