A Low-Budget Approach for Vehicle Detection and Occlusion Removal on Traffic Videos

Abstract

A huge increase in traffic flow in large cities has created serious problems for urban planning and transit authorities, as they seek for solutions for handling increasingly larger traffic jams. In this paper, we describe part of an automated system designed to extract information from videos captured by closed-circuit surveillance camera systems to perform traffic monitoring. The main contributions of this paper are a combination of known methods for performing robust vehicle detection and counting as well as the adaptation of the Viola and Jones framework to deal with occlusion events. We tested our system on videos acquired at different times of the day and the results show a dear improvement over the traditional Viola and Jones method. We also compared the results of two versions of our approach to YOLO V3, a state-of-the-art method for detecting and classifying objects, on three different computers, and achieved slightly lower occlusion solving rates. We emphasize that YOLO V3 worked at very low FPS rates for one computer when not using the GPU, when compared to the Frames Per Second (FPS) rates achieved by our approach. Moreover, YOLO could not be executed on the notebook used in the experiments or a low-budget device Raspberry Pi 2, while our approaches achieved processing rates between 8 and 10 FPS on the Raspberry Pi 2 model.