适用于硬件高速计算的CNN目标跟踪算法  

A Target Tracking Algorithm Based on CNN for High-speed Calculation of Hardware

作　　者：季飞 王加庆 刘剑[1,2] 吴南健 JI Fei;WANG Jia-qing;LIU Jian;WU Nan-jian(State Key Laboratory of Super Lattices and Microstructures,Institute of Semiconductors,Chinese Academy of Sciences,Beijing 106083,China;University of Chinese Academy of Sciences,Beijing 10008a,China)

机构地区：[1]中国科学院半导体研究所超晶格国家重点实验室,北京100083 [2]中国科学院大学,北京100083

出　　处：《微电子学与计算机》2018年第12期115-118,124,共5页Microelectronics & Computer

基　　金：国家自然科学基金(61234003;61434004;61504141);北京脑计划项目(z161100000216129);中国科学院学科项目(kjzd-ew-111-04)

摘　　要：针对硬件上实现指定目标的高速跟踪,提出一种适用于硬件高速计算的深度卷积神经网络(convolutional neural network,CNN)目标跟踪算法.通过分析卷积层、卷积核、亚采样层和激活层对于网络性能的影响,针对硬件实现构建多种CNN结构.训练指定目标样本,得到基于卷积深度特征的目标模型,采用灵活的搜索策略,调用优化后的模型参数实现硬件上的目标跟踪.结合实例,对比了多种网络的性能和跟踪效果,其中最优模型参数仅为368Byte,测试错误率为0.0125,跟踪误差均值为0.779像素,证明了该算法在硬件上实现目标追踪的有效行和可行性..To achieve the specified target tracking on the hardware.A target tracking algorithm based on convolution neural network for hardware is proposed.By analyzing the effects of the coiling layer,convolution kernel, subsampling layer,and activation layer on network performance,various CNN structures are constructed for hardware implementation.By training the target sample,the target model based on the convolution depth characteristics is obtained.The optimized model parameters are invoked to track the target tracking on the hardware,by using a flexible search strategy.The performance and tracking effect of a variety of networks are compared with an example.The parameter size of the optimal model is 368Byte,and the test error rate is 0.0125, and the mean of tracking error is 0.779.The effectiveness and feasibility of the algorithm to achieve target tracking on hardware is proved.

关 键 词：卷积神经网络 硬件计算 目标跟踪 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]