【作者】 徐敏；

【导师】 王天江；

【作者基本信息】 华中科技大学 ， 计算机应用技术， 2020， 硕士

【摘要】 尽管深度神经网络目前在许多方面展示出强大的能力,但由于其高能耗的特性,无法应用在一些计算资源有限的场景下。而脉冲神经网络作为第三代神经网络,是目前最拟合生物神经元机制的模型。相比深度神经网络,在脉冲神经网络所需的计算量和能耗远小得多。目前,脉冲神经网络在图像识别和目标检测方面达到了较好的效果,但对于更复杂的计算机视觉任务相关研究却很少。作为计算机视觉领域最重要的任务之一,目标跟踪有较高的实时性和精确度要求,探索脉冲神经网络在目标跟踪中的应用具有极大的研究意义。为了优化脉冲序列的时间分布,在目标跟踪模型中降低时延,研究提出了一种双电压阈值编码机制,该机制将脉冲序列限制在周期性的固定分布中以增加脉冲密度。目前主流跟踪模型采用了基于匹配的孪生神经网络,研究提出了基于膜电位和时间序列的简单匹配方法,以及适用于目标跟踪的脉冲混合相似性匹配法,更好地利用了脉冲中的时序信息来准确计算模板图像和搜索图像之间的响应图。基于双电压阈值编码机制以及脉冲混合相似性匹配法,研究提出了第一个可用于复杂场景实时目标跟踪的全卷积脉冲孪生网络Spiking-SiamFC。在OTB-2013,OTB-2015和VOT2016数据集上的实验结果证明,Spiking-SiamFC达到了与SiamFC近似的精度且具有较低的时延,可以在神经形态芯片TrueNorth上达到实时跟踪效果。更多还原

【Abstract】 Although deep neural networks(DNNs)have shown the great power in many scenarios,they can not be employed in systems with limited computing resources due to their heavy energy consumption.As the third generation of neural networks,spiking neural networks(SNNs)best fit biological neuron mechanism.Compared with DNNs,the calculation and energy consumption of SNNs are much smaller.Now SNNs have shown the effectiveness in image recognition and object detection,but there are few researches on more complex computer vision tasks.As one of the most important tasks in computer vision,object tracking has high real-time and accuracy requirements.It is of great significance to explore the application of SNNs in object tracking.In order to optimize the time distribution of the spike train and reduce the time delay in tracking models,a two-status coding scheme is proposed.Now the mainstream of tracking model adopts Siamese structure based on matching mechanism.Therefore Potential Spiking Matching,Temporal Similarity Matching and Hybrid Spiking Similarity Matching are proposed.They fully use time sequence information to accurately calculate the similarity between the template image and the search image.Based on Two-Status Coding mechanism and Hybrid Spiking Similarity Matching method,Spiking-SiamFC is proposed.It is the first fully-convolutional siamese spiking network which can be used for real-time object tracking in complex scenarios.Experimental results on OTB-2013,OTB-2015 and VOT2016 data sets show that Spiking-SiamFC achieves similar accuracy to SiamFC and has a low time delay,which can achieve real-time tracking effect on the neural morphological chip like TrueNoth.更多还原