一种改进的 SSD 网络车载图像检测方法-《计算机技术与发展》

文章信息/Info

Author(s):: BAO Run-jia; HOU Qing-shan; XING Jin-sheng; School of Mathematics and Computer Science,Shanxi Normal University,Linfen 041000,China

Keywords:: SSD model; detection accuracy; convolutional neural network; object detection; feature layers fusion

摘要:: 对行人和车辆的检测识别是无人驾驶领域的重要组成部分,为满足该领域对相关模型检测精确度的需求,以传统单发多框检测器(single shot multibox detector,SSD)为基础,提出了一种车载图像识别改进算法。鉴于传统 SSD 目标检测算法不能充分利用局部特征和全局语义特征、目标定位和识别存在矛盾等缺陷,提出了 SSD 检测模型相关特征层的融合方法,从而重新生成模型的目标检测金字塔(object detection pyramid,ODP)。改进算法将输入图像中待检测目标的低层次细节特征与高层次语义特征结合起来,降低了待检测目标定位与识别间的矛盾,达到了提升模型检测精确度的目的。利用行车记录仪获得的车载图像数据集进行训练,实验结果表明,改进的 SSD 算法在相关图像数据集的测试集上可以达到79.2% 的精确度,与传统的 SSD 算法相比精确度提高了 2.3%。

Abstract:: Detection and recognition of pedestrians and vehicles is an important part of the unmanned driving field. In order to meet the demand of dete-ction accuracy of the algorithm model in this field,an improved algorithm for vehicle image recognition is proposed based on the traditional SSD (single shot multibox detector) network. Since the traditional SSD object detection algorithm cannot make full use of local features and global semantic features,and there are contradictions in object location and recognition, a fusion method of the relevant feature layers of the SSD detection model is proposed to regenerate the ODP (object detection pyramid) of the model. The improved algorithm combines the low-level detail features and high-level semantic features of the objects to be detected in the input image,reducing the contradiction between the positioning and recognition of the objects to be detected,and achieving the purpose of improving the model detection accuracy. The vehicle images data set obtai-ned by the driving recorder is used for training. The experiment shows that the improved SSD algorithm can achieve 79.2% accuracy in the test set of related images data set, and the accuracy is improved by 2.3% compared with the traditional SSD algorithm.