基于卷积神经网络的行人目标检测系统设计

Design of Pedestrian Target Detection System Based on Convolutional Neural Network

为获得最直观的行人目标检测结果,避免运动姿态不确定性对实时检测造成的影响,设计基于卷积神经网络的行人目标检测系统;以CNN计算框架作为硬件结构主体,分级连接目标传感器与神经型卷积分类器,按照并行检测原理及卷积神经架构搭建检测体系结构;建立训练文件体系,通过迎合目标训练环境的方式,配置必要的检测文件参数,完成待检测行人目标的样本训练处理;在检测节点架构中,规定与访问接口关联的配置条件,借助增设的模块复用加速结构,直接获取行人目标检测结果,实现行人目标的样本重构,完成基于卷积神经网络的行人目标检测系统设计;实验结果表明,与PCA、SVM算法相比,应用卷积神经网络型检测系统后,单位时间内的行人目标检测量达到9.6×10^9 T,目标数据堆积速率降低至1.14×10^9 T/s,能够直观获取行人目标检测结果,有效抑制了运动姿态不确定性对系统实时检测的影响。

In order to obtain the most intuitive pedestrian target detection results and avoid the impact of motion pose uncertainty on real-time detection,apedestrian target detection system based on convolutional neural network is designed.The CNN computing framework is used as the main body of the hardware structure,the target sensor and the neural type convolutional classifier are connected in a hierarchical manner,and the detection architecture is built according to the parallel detection principle and the convolutional neural architecture.Establish a training file system,configure the necessary detection file parameters by catering to the target training environment,and complete the sample training process of the pedestrian target to be detected.In the detection node architecture,the configuration conditions associated with the access interface are specified,and the additional module multiplexing acceleration structure is used to directly obtain the pedestrian target detection results,and the pedestrian target sample reconstruction is realized,and the pedestrian target detection system based on the convolutional neural network is completed design.Experimental results show that,compared with PCA and SVM algorithms,after applying the convolutional neural network detection system,the pedestrian target detection amount per unit time reaches 9.6×10^9 T,and the target data accumulation rate is reduced to 1.14×10^9 T/s,which can be obtained intuitively Pedestrian target detection results effectively suppress the impact of motion pose uncertainty on real-time detection of the system.