针对红外图像行人检测任务中图像细节少、特征信息提取困难、检测准确率低等问题, 提出一种改进的SSD红外行人检测算法。利用深度可分卷积方法降低特征提取网络参数数量和网络运算量, 改善实时性; 在网络中嵌入SENet模块, 重新分配各特征通道权重, 提升网络对行人目标针对性; 针对行人目标空间占比固定的特点, 通过聚类分析算法设定Default Boxes数值, 提升行人检测效果。实验结果表明, 所提改进算法优于VSSD算法, 查准率和查全率分别达到91.7%和84.8%, 同时,算法实时性也得到大幅改善。

To solve such problems in infrared image pedestrian detection as lack of image details, low detection accuracy and the difficulty in extracting the feature information, an infrared image pedestrian detection algorithm based on the improved SSD algorithm was proposed.Depthwise Separable Convolution (DSC) method was used to reduce the number of parameters and calculation costs of the network in feature extraction, and improve the real-time performance.The SENet module was embedded in the network, and the weight of each feature channel was reassigned, so that the network could better target the pedestrians. Since the space ratio of the pedestrian target is relatively stable, the Default Boxes value was set by using cluster analysis algorithm, so as to improve pedestrian detection effects.Experimental results showed that, the proposed algorithm performed better than the VSSD algorithm.The precision ratio and the recall ratio reached 91.7% and 84.8% respectively, and the real-time performance of the algorithm was also greatly improved.