Abstract: In order to obtain an initial candidate frame that conforms to the morphological characteristics of vehicles more accurately, a vehicle detection algorithm based on the improved Faster R-CNN model is proposed. First, the coordinate values of target frames are extracted to get width and height values of labeled boxes, and then K-means algorithm is used to cluster the width and height values of all boxes. Second, by resetting the anchor box size and the anchor box ratio of region proposal network (RPN) according to the coordinates of the cluster center point, the three sizes and three ratios of the Faster R-CNN can be improved. Finally, vehicle data of four types including cars, SUVs, buses and trucks are selected to train both the unimproved and the improved Faster R-CNN models. At the same time, the performance of the two models in vehicle detection and vehicle identification tasks are compared. Experimental results show that the improved Faster R-CNN model can achieve 84.69% detection accuracy, which is 3.12% higher than the original model. The algorithm effectively improves the missed detection and false detection problems, and shows high robustness in bad weather and complex background.

Key words: vehicle detection, Faster R-CNN, K-means clustering