A Convenient Approach for Lane-Level Congestion Detection with On-Board Camera Images and Vehicle Data

Lane-level traffic congestion detection is crucial for intelligent transportation systems to provide drivers of precise and timely traffic information. However, traditional traffic monitoring methods relying on fixed sensors or GNSS data face limitations in terms of cost and accuracy at the lane-level. In this paper, we propose a convenient approach using on-board camera images and vehicle data for lane-level congestion detection. We leverage an object-graph approach to incorporate spatial and temporal dynamics of vehicles in sequential scenes to accurately grasp their lane positions under severe occlusions, and ego-vehicle's speed information to estimate traffic flow. To demonstrate the effectiveness of our proposed approach, we collect a large-scale real dataset of on-board camera images and vehicle data from connected vehicles driving on roads, and conduct an evaluation against several baseline models. The proposed algorithm outperformed the baselines by 0.02 (equivalently, 2.9%) in terms of the weighted average F1 score across all lanes. This study provides an important foundation for further research on the lane-level traffic congestion detection by identifying the key challenges and insights through our experiments.