改进YOLOv1的视频图像运动目标检测

Moving target detection in video images based on improved YOLOv1 network

针对视频图像中运动目标位置和大小变化频繁的特点,通过改进网络结构和训练过程,搭建了基于YOLOv1的神经网络框架。该框架采用ResNet50进行特征提取,增加卷积层和全连接层优化对不同尺度特征信息的传递,通过Sigmoid层和BN层在稳定输出结果的同时,加快训练速度。PASCAL VOC2007数据集和实景视频数据的实验表明,相比原始YOLOv1,本文方法的FPS和mAP分别提高了4.44%和4.57%,满足视频图像运动目标检测的实时性和精度要求。

By improving the network structure and training process,a neural network based on YOLOv1 framework is present to deal with the frequent position and size changes of moving targets in video images. Particularly,ResNet50 is utilized for feature extraction,convolution layer and full connection layer are appended to optimize the transmission of feature information at different scales,and Sigmoid and BN layer are employed to speed up the training speed while stabilizing the output results. Experimental results of PASCAL VOC2007 and self-captured video dataset show that the FPS and mAP of the method proposed in this paper are improved by 4.44% and 4.57% respectively compared with the original YOLOv1, which can meet the requirements of real-time and accuracy of moving target detection in video im ages.