【摘要】 针对YOLOv4-tiny网络特征提取能力弱，多层特征融合不足导致的检测精度低的问题，论文提出了一种基于YOLOv4-tiny的交通目标检测模型（SS-YOLOv4-tiny）。首先，往特征提取网络中引入深度可分离卷积以及空间深度可分离卷积融合模块，在降低网络参数量和计算量的同时增强特征提取能力；然后，提出自适应合成金字塔特征融合网络代替原有的单向特征融合网络，使得不同层次的特征图能进行更平滑且充分的信息交流。在KITTI数据集上的实验结果表明，改进后的模型在测试集上的平均检测精度相比原YOLOv4-tiny网络提升了4.9%，具有较好的检测精度和实时性，适合部署在车载服务器等边缘设备上。更多还原

【Abstract】 To address the problem of low detection accuracy due to weak feature extraction ability and insufficient multi-layer feature fusion in YOLOv4-tiny network,a traffic target detection model based on YOLOv4-tiny(SS-YOLOv4-tiny)is proposed in this paper. Firstly,a depth-separable convolution and a spatial depth-separable convolution fusion module are introduced into the feature extraction network to enhance the feature extraction capability while reducing the number of network parameters and computational effort. Secondly,the adaptive synthetic pyramidal feature fusion network is proposed to replace the original unidirectional feature fusion network,which enables a smoother and sufficient information exchange between different levels of feature maps. Experimental results on the KITTI dataset show that the average detection accuracy of the improved model on the test set is improved by 4.9% compared with the original YOLOv4-tiny network,which has better detection accuracy and real-time performance and is more suitable for deployment on edge devices such as in-vehicle servers.更多还原