Accurate Vehicle Counting Approach Based on Deep Neural Networks

Abstract

Vehicle counting is considered one of the most important applications in traffic control and management. To count vehicles, synchronous vehicle detection and tracking should be carried out. Recently, detection via deep neural networks (DNN) has achieved good performance. However, exploiting the DNN efficiently for vehicle counting is still challenging. In this paper, an efficient approach for vehicle counting employing DNN and KLT tracker is proposed. To decrease the time complexity, vehicles are detected via DNN every N-frames, N=15 for example. Trajectories are extracted by tracking corner points through the N-frames. Then an efficient algorithm is introduced to assign unique vehicle labels to their corresponding trajectories. The proposed results, performed on diverse vehicle videos, show that vehicles are accurately tracked and counted whatever they are detected one or more times by the DNN.