Target detection is an important research direction in the field of computer vision. Although the single-shot detector (SSD) model achieves good results in terms of detection accuracy and speed, its use of shallow features with low semantic information for training small targets is prone to target misses and false detections. In this paper, an improved SSD target detection model based on a once bidirectional feature pyramid network (OBSSD) is proposed. First, a bidirectional feature fusion module is constructed based on the principle of hierarchical fusion to solve the problem of under-utilization of shallow features. Second, fusion weights are introduced to fuse features at different levels more effectively and mitigate the problem of low semantic information of shallow features. Finally, a detection unit based on the residual module is added before classification and regression prediction to address the problem of inaccurate target localization caused by the biased translational invariance of the classification network. The experimental results on the PASCAL VOC2007 test set show that the mean average precision (mAP) of the proposed model is 80.8%, which is 6.5 percentage points higher than that of the SSD model, and the detection speed meets the demand for real-time detection.