engleski

Experimental Evaluation of Vehicle Detection Based on Background Modelling in Daytime and Night-Time Video

Vision-based detection of vehicles at urban intersections is an interesting alternative to commonly applied hardware solutions such as inductive loops. The standard approach to that problem is based on a background model consisting of independent per-pixel Gaussian mixtures. However, there are several notable shortcomings of that approach, including large computational complexity, blending of stopped vehicles with background and sensitivity to changes in image acquisition parameters (gain, exposure). We address these problems by proposing the following three improvements: (i) dispersed and delayed background modeling, (ii) modeling patch gradient distributions instead of absolute values of individual pixels, and (iii) significant speed-up through use of integral images. We present a detailed performance comparison on a realistic dataset with handcrafted groundtruth information. The obtained results indicate that significant gains with respect to the standard approach can be obtained both in performance and computational speed. Experiments suggest that the proposed combined technique would enable robust real-time performance on a low-cost embedded computer.

background modelling ; vehicle detection ; virtual inductive loops

nije evidentirano