Moving Vehicle Detection Combining Edge Detection and Gaussian Mixture Models

Abstract

For the computationally intensive problems of the traditional Gaussian mixture model method and “ghosting” when transforming stationary objects into motion for a long time, a moving vehicle detection method that combines difference of Gaussian (DoG) edge detection with an adaptive Gaussian mixture model (Adaptive GMM) method is proposed. First, the traffic surveillance video is pre-processed to reduce picture redundancy and improve picture quality through compression and smoothing. Then, DoG edge detection was applied to each frame to extract the contour of the moving vehicle, and a modified multi-frame averaging method was used to obtain a pure background image for the first N frames, and an adaptive Gaussian mixture model method was applied to update the background model in real-time by judging whether the new pixels match the existing model to decide whether to perform parameter updates. Finally, filling “holes” in detected moving vehicles through post-processing methods to make the vehicle detection results clearer and more complete. Through experimental analysis, the results show that the detection algorithm average accuracy reaches 96%, which is 2.85% higher than the original GMM, and the average time spent per frame reaches 18.6 ms, which meets the accuracy and real-time requirements of the intelligent transportation system.