Traffic Scene Object Detection Algorithm Based on Improved SSD

With the emergence of electric vehicles and the development of autonomous driving technology, natural scene traffic sign detection has become a hot research topic. Although traffic sign detection has made great progress, there are still some problems that need to be solved: 1. traffic signs have large-scale variations, and it is difficult to capture the features of targets with large-scale variations; 2. traffic signs have more small targets, and the information of small targets is less to make detection difficult. In order to solve the above problems, we propose a spatial aggregation feature enhancement network based on improved SSD(RSC-SSD), aimed at extracting richer object detail information Specifically, we firstly apply the channel attention mechanism to the backbone network to improve the target extraction capability. Secondly, we propose a Feature Aggregation Module (CSPP) to enhance the background information, which can effectively capture the features of small targets mainly consisting of two parts: the dilated convolution of different sizes and the spatial attention mechanism. Then, we propose a learnable feature fusion module(FFM) to reduce information loss when adjacent features are fused. Finally, we conducted experiments on the CCTSDB dataset, and compared it with the baseline model SSD network, the detection accuracy of the model in this paper is improved by 4.4%, which fully demonstrates that the algorithm in this paper can be well adapted to traffic scene detection.