Due to the complex environment of the target in road traffic, there exist the problems of the insufficient extraction of key features by the model and the low accuracy of target positioning. The SSD model is used as the basic framework in this paper, and research is conducted on feature extraction methods, key information enhancement, and non-local feature positioning. Firstly, in order to solve the multi-scale problem of targets in road traffic scenarios, a jumping reverse feature pyramid structure is proposed to generate more discriminant features. Secondly, in order to solve the problem that information at different semantic levels has different degrees of contribution to the feature fusion process, an adaptive feature fusion module based on attention mechanism is designed to enhance the key feature expression ability non-priori at the channel level. Finally, the cross-attention module is introduced to improve the position sensitivity of the model to the target. Experimental results indicate that compared with the original model of SSD, in guarantee under the condition of real-time, the average accuracy of the proposed algorithm is improved by 2.6% on PASCAL VOC sub-dataset and 3.9% on homemade road traffic dataset. Taking everything into account, the improved algorithm can be applied widely to the task of detecting vehicles and pedestrians on the road.