With the rapid growth of the number of motor vehicles in China, inevitably there would appear a series of severe problems concerning safety and traffics. At the same time, the video image files are increasing at an explosive speed, which has brought a lot of trouble to the public security monitoring, criminal investigation and the case detection. It is important to research an efficient and accurate vehicle detection algorithm. This paper proposes a new deep convolution neural networks frame for vehicle detection and coarse grained recognition based YOLO method. Multilayer perceptron convolution layers are added in the new network structure framework to enhance nonlinear ability of feature mapping. This framework deletes fully connected layers and predicts the bounding boxes using anchor boxes. The new framework improves recall rates of object detection and effectively reduces computational complexity. Experimental results show that the improved method has an average accuracy of 94.7% for vehicle detection under iteration 20000 times. Compared with other detection methods, the processing speed and accuracy of the new method have been improved.