一种卷积神经网络的车辆和行人检测算法

Vehicle and pedestrian detection algorithm based on convolutional neural network

针对传统车辆行人检测算法速度慢、精度低、难以进行多尺寸检测等问题,设计了一种深度卷积神经网络,应用到密集车辆和行人的实时检测。采用批标准化、Res结构、SSES结构搭建深度卷积神经网络,融合不同深度的特征并用3个不同尺寸特征图检测以提高网络对多尺寸目标的检测能力。神经网络在PASCAL VOC2007、2012车辆和人混合数据集上训练,检测车辆、行人的平均精度分别到达80.1%、83.6%,检测单张图片耗时0.04 s。实验结果表明,设计的算法检测车辆、行人精度较高,多尺寸目标检测效果好,可实时检测。

Aiming at the problems of slow speed,low precision and difficulty in multi-size detection of traditional vehicle pedestrian detection algorithms,a deep convolutional neural network is designed and applied to real-time detection of dense vehicles and pedestrians.The deep convolutional neural network is built by batch normalization,residual networks and split-squeeze-excitation-and-sum networks.It combines the features of different depths and detects multi-size objects with three different size feature maps to improve network detection ability.Neural network is trained on PASCAL VOC 2007 and 2012 car and person mixed data sets,the average accuracy of car and person detection reaches 80.1%and 83.6%respectively,the detection of a single picture takes 0.04 s.The experimental results show that the proposed algorithm has high accuracy in detecting car and person,good results in multi-size object detection,and real-time detection.