With the continuous development of intelligent transportation, the key of intelligent transportation is forward vehicle location. Accurate and reliable forward vehicle location results are of great significance to the safe driving of intelligent vehicles. In order to solve the problem of forward vehicle location and tracking, this paper proposes a method of license plate feature point detection based on convolutional neural network. Firstly, according to the prior size information of the license plate, the efficiency of license plate detection is improved by modifying the prior frame size of Tiny-YOLOv3 and improving the network structure of Tiny-YOLO-v3. Secondly, based on the structure of the residual neural network, a residual neural network was designed to detect the feature points of the license plate with a fixed aspect ratio. In this model, the anchor frame-free object detection method is applied to license plate detection. Instead of using anchor frame to obtain proposal license plate regions, the license plate corner is predicted based on the heat map, According to the geometric structure of the license plate, a loss function containing the relative position information of the feature points was designed to realize the accurate regression of the feature points of the license plate. On the one hand, the performance of the improved YOLO algorithm based on CCPD public license plate data set, the recognition accuracy can reach 78.3%, and the average processing time is 178ms. On the other hand, the license plate image data set was obtained based on YOLO algorithm, and the validity of the feature point detection model based on residual neural network was evaluated. The mean square error of the license plate feature point regression model in the license plate image test set was 0.036, and the average processing time was 185ms. Experimental results on public license plate data sets show that the proposed method has high accuracy in detecting the location of license plate feature points, and the relative positions of the detected feature points remain unchanged. The effectiveness of the proposed method is verified.