Most of the current vehicle detection models based on depth learning have many problems such as large parameters, high computational complexity, and most lightweight target detection models can not meet the actual work requirements in accuracy. Aiming at this problem, this paper proposes a vehicle detection method based on improved YOLOv3. In order to reduce model parameters and speed up network inference, CSPNet and Ghost modules are added to the Backbone part of YOLOv3; Expand the feature input size of the feature fusion network, add the interval fusion structure of four times of up sampling on the basis of the original two times of up sampling, and use Depth Separable Convolution instead of conventional convolution to further compress the model parameters. The experimental results show that the accuracy of the improved model is 1.5% higher than that before the improvement, the detection speed is 23.0ms faster, and the model has good robustness, which can meet the detection requirements of different scenes.