The mainstream target detection algorithms in the field of automatic driving have poor detection effect on small targets, which poses a threat to driving safety. The one-stage anchor-free CenterNet algorithm is improved to solve this problem. Firstly, the original backbone network is replaced by ResNeSt50 network with split-attention mechanism, and the ReLU activation function is upgraded to FReLU, which strengthens the effect of feature extraction with little additional computational overhead; Then, a lightweight network PASN is proposed to fuse semantic features of different scales, and Spatial Pooling Pyramid (SPP) module is introduced into the shallow feature input to enhance the expression of small target information; Finally, random multi-scale input training is carried out on Kitti data set. The verification set results shows that the FPS of the improved algorithm reaches 37.7, meets the real-time requirements, the average precision of small targets is improved by 12.9% and the mean average precision is improved by 13.9%， At the same time, the detection speed and average precision are higher than the mainstream algorithm Yolov4；It can detect 31 images per second on the real vehicle, which provides strong support for the development of automatic driving technology and has engineering application value.