EFFICIENT APPROACH FOR VEHICLE COUNTING BASED ON DEEP CONVOLUTIONAL NEURAL NETWORKS

摘要

Vehicle counting plays an important role in intelligent transport systems. A vehicle counting system should be fast to be implemented in real-time circumstances. Recent methods for vehicle counting usually include two stages, vehicle detection and vehicle tracking. Vehicle tracking stage requires more computational cost, which makes the system less efficient. In this paper, a new approach for vehicle counting based on deep convolutional neural networks (CNN) is proposed. First, an improved single shot multibox detector (SSD) is proposed for fast vehicle detection. The base network and detection network in the original SSD are replaced and modified to reduce the computational cost. To eliminate vehicle tracking step, a region of interest (ROI) is set in each image frame. The number of vehicles is increased when a vehicle is passing the ROI. Furthermore, an improved algorithm for counting exactly the number of vehicles is introduced. Experimental results on public datasets show that the proposed method is the fastest system for counting vehicle among current systems and achieves comparable accuracy compared with state-of-the-art methods.