An enhanced three-frame-differencing approach for vehicle detection under challenging environmental conditions

Abstract

Robust and efficient vehicle detection, counting and tracking is an important task in Intelligent Transportation Systems. With the continuous development of computer vision technologies, remarkable progress has been made in vision-based vehicle detection. Comparing to other sensors, vision cameras provide rich information for driving understanding. At the same time, robust feature descriptors and efficient background models have been proposed for the purpose of accurate vehicle detection. In this thesis, a computationally efficient method for vehicle detection, counting and tracking under different environmental conditions is presented, with a special focus on adverse illumination and weather. The general framework is based on enhanced Three-Frame-Differencing (E-TFD). In a given video sequence, three consecutive frames are utilized to generate frame differencing images. With an efficient thresholding and removal of small noise regions, moving vehicles can be extracted in an efficient and accurate manner. Meanwhile, based on extracted regions of interest (ROIs), exact numbers of vehicles can be counted and displayed on the screen. The E-TFD method can detect and count vehicle candidates in both fine and inclement weather conditions, including sunny, rainy, foggy, snowy, blizzard, wet snow and nighttime conditions in this study. To evaluate the E-TFD detection approach, nine videos are collected from different sources. Six videos are selected from two public datasets, CDnet 2014 and KIT dataset, for the purpose of performance analysis. At the same time, three videos are recorded from different roads in Kowloon, Hong Kong using a digital camera. Of all 4532 tested frames, 10059 vehicles can be successfully detected out of 11556 vehicles, showing an average detection rate of 87.1%. The E-FTD method shows a significant improvement of detection rate in adverse conditions and can provide a efficient solution of all-time, all-weather detection, counting and tracking that can in future be embedded into a real-time traffic surveillance system.