To improve the accuracy of the algorithm for vehicle detection and solve the problem that the original algorithm is not effective in the complex traffic scene, a vehicle detection method based on attention mechanism and improved densely connection network structure was proposed. Firstly, SoftPool was used in the transition layer to consolidate the characteristic information between the dense blocks. Secondly, the expression of effective channel features was enhanced by the lightweight channel attention mechanism, it was used as the deep feature extraction layer of Darknet-53. The CIOU loss was used as the prediction loss term of the bounding box position of the model, and reduce the model volume using deep separable convolution. Compared with the original algorithm, the mAP value is increased by 2. 6%, and the model volume is reduced to 42%. Experimental results show that the algorithm has good detection performance in complex traffic scene.