Multiple Hypothesis Detection and Tracking Using Deep Learning for Video Traffic Surveillance

Abstract

Moroccan Intelligent Transport System is the first Moroccan system that uses the latest advances in computer vision, machine learning and deep learning techniques to manage Moroccan traffic and road violations. In this paper, we propose a fully automatic approach to Multiple Hypothesis Detection and Tracking (MHDT) for video traffic surveillance. The proposed framework combines Kalman filter and data association-based tracking methods using YOLO detection approach, to robustly track vehicles in complex traffic surveillance scenes. Experimental results demonstrate that the proposed approach is robust to detect and track the trajectory of the vehicles in different situations such as scale variation, stopped vehicles, rotation, varying illumination and occlusion. The proposed approach shows a competitive results (detection: 94.10% accuracy, tracking: 92.50% accuracy) compared to the state-of-the-art approaches.

Keywords

• Traffic surveillance
• computer vision
• deep learning
• Kalman filter
• data association
• detection