Traffic Congestion Estimation Using HMM Models Without Vehicle Tracking

We propose an unsupervised, low-latency traffic congestion estimation algorithm that operates on the MPEG video data. We extract congestion features directly in the compressed domain, and employ Gaussian Mixture Hidden Markov Models (GM-HMM) to detect traffic condition. First, we construct a multi-dimensional feature vector from the parsed DCT coefficients and motion vectors. Then, we train a set of left-to-right HMM chains corresponding to five traffic patterns (empty, open flow, mild congestion, heavy congestion, and stopped), and use a Maximum Likelihood (ML) criterion to determine the state from the outputs of the separate HMM chains. We calculate a confidence score to assess the reliability of the detection results. The proposed method is computationally efficient and modular. Our tests prove that the feature vector is invariant to different illumination conditions, e.g. sunny, cloudy, dark. Furthermore, we do not need to impose different models for different camera setups, thus we significantly reduce the system initialization workload and improve its adaptability. Experimental results show that the precision rate of the presented algorithm is very high around 95%.