Abstract: An effective approach for visual traffic data collection based on multi-features fusion is presented. The main idea is to configure several virtual loops (detection zones) on the image, assuming moving vehicles may cause pixel intensities to change, then by identifying such pixel changes, to detect vehicles and estimate vehicle speed. The contributions of this paper are: 1) We integrate multiple features including foreground area, texture change, and pixel motion in the virtual loop to detect vehicles, and present an effective multi-features fusion approach, which can significantly improve the accuracy of vehicle detection; 2) we search the pixel motion vectors in a single virtual loop to estimate vehicle speed, avoiding the mismatching problem existing in the dual-loop detector. Algorithmic testing results show that the proposed approach is able to accurately detect vehicles and estimate vehicle speed in a wide range of traffic scenes and weather conditions. Based on the proposed approach, we developed an embedded traffic video detector, which can operate at traffic intersections to collect traffic data and provide decision-making basis for traffic signal control and traffic law analysis.