Luminance Adaptive Dynamic Background Models for Vision-Based Traffic Detection

Measuring traffic flow by employing vision-based detection suffers from several challenges, particularly the illumination variation. Consequently, this research focuses on solving traffic detection problem due to both sudden and gradual illumination changes. A number of theories are proposed to define different components of an image. Specifically, first-order model for illumination variation and Fourier series for incorporating traffic arrival patterns are considered to define background and foreground, respectively. We have utilized these definitions to formulate the traffic detection problem and subsequently three adaptive dynamic background models have been developed to solve it. The third model that incorporates both luminance and pollution controlling parameters fixes the problems and limitations faced by the first and second models. Besides, a new per pixel binary threshold model related to the third model is also developed for foreground segmentation. Using a real video dataset, a constrained optimization is performed to determine the optimal values of model parameters, where the feasible regions of the parameters are obtained graphically. The model validation using a separate video dataset shows more than 95% Percent Correct Classification (PCC) value and around 90% Precision and Recall values. Additionally, a field test is conducted in three different locations and the performance of the model is evaluated. Evaluation shows that, the model achieves the highest value of 93% in terms of Average Accuracy of Object Count (AAOC) for urban arterial dataset, which represents its robustness in object detection.