Detailed evaluation and analysis of vision-based online traffic parameters estimation approach using low resolution web cameras

In this paper, we give an overview and a detail analysis of our approach for vision-based real-time traffic parameters estimation using low-resolution web cameras. Traffic parameters estimation approach mainly includes three major steps, (1) stable background estimation, (2) vehicle detection, mean speed and traffic flow estimation, and (3) traffic scene classification into three states (normal and congested). The background image is estimated and updated in realtime by novel background estimation algorithm based on the median of First-in-First-Out (FIFO) buffer of rectified traffic images. Vehicles are detected by background subtraction followed by post-processing steps. By exploiting the domain knowledge of real-world traffic flow patterns, mean speed and traffic flow can be estimated reliably and accurately. Naive Bayes classifier with statistical features is used for traffic scene classification. The traffic parameter estimation approach is tested and evaluated at the German Aerospace Center’s (DLR) urban road research laboratory in Berlin for 24 hours of live streaming data from web-cameras with frames per second 1, 5 and 10. Image resolution is 348 x 259 and JPEG compression is 50%. Processed traffic data is cross-checked with synchronized induction loop data. Detailed evaluation and analysis shows high accuracy and robustness of traffic parameters estimation approach using low-resolution web-cameras under challenging traffic conditions.