OSTM-NET: Joint scale variation and occlusion handling deep network for real-time vehicle counting and volume estimation

Abstract

Vehicle counting and traffic volume estimation using videos are difficult tasks crucial for efficient traffic control in smart cities. Several existing techniques rely in tracking and detecting mechanisms of the vehicles. These methods are ineffective for detecting occluded and small vehicles. Also, contextual information loss occurs in deep counting networks. An innovative Internet-of-Things (IoT)- driven Intelligent Transportation Management (ITM) system is proposed to address these issues. Initially, traffic videos are converted to temporal-spatial images instead of using complex detection and tracking methods. The density map for the temporal spatial image is estimated using an occlusion-aware spatio-temporal multi-scale network (OSTM-Net). It consists of two sub-networks for capturing occluded and small vehicles simultaneously. The scale-aware column network (SCNet) accurately captures small vehicles and preserves contextual information through enhanced scale representation. At the same time, the occlusion management network (OM-Net) uses position-sensitive regions of interest (PSRoI) deformable pooling to address the occlusion issues. Finally, volume estimation and counting are calculated in accordance with the density map obtained from OSTM-Net. Every path in the videos is processed separately using OSTM-Net to calculate the vehicle count in every path for effective traffic control in this proposed approach. Furthermore, the effectiveness of the sub-networks (SCNet and OM-Net) is validated using ablation experiments. The proposed IoT-based ITM achieves high performance in counting vehicles and estimating traffic volume compared to other existing approaches.