Accelerated Fog Removal from Real Images for Car Detection

Image de-hazing improves the visual quality of images in computer vision applications, such as object detection and object tracking. Fog removal from car photos taken by street cameras is considered essential to accurate car detection. An accelerated image enhancement technique is presented for car detection as part of an effort to count cars using existing street cameras for the purpose of traffic management. Two aspects of car detection are tackled: 1) An existing image fog removal technique is accelerated by replacing a time consuming image filter with a faster filter while maintaining negligible image degradation, 2) A quick and practical algorithm to detect a car in a fog-free image is proposed and applied to a database of about 100 car images. The main idea is to give an indication of the capacity of cars on a given road by counting them using street cameras in the presence of fog. Such car counting method can assist traffic centers to manage traffic flow and prevent traffic incidents. The devastating effect of fog-related accidents inspired this research to develop a fast execution-time algorithm to detect cars in the presence of heavy fog using existing road cameras. Acceleration is the main goal of this research, in addition to car detection accuracy. In order to achieve the required acceleration and accuracy, several image processing techniques are investigated. The techniques are proposed to accelerate fog removal from car images and accurately detect cars with an execution time, faster than any other existing fog removal and car detection technique. Therefore, the developed techniques provide a viable solution to a difficult problem in the area of intelligent transportation systems. The improved fog removal technique is performed by estimating the transmission map using the Proposed Adaptive Filter (PAF) to recover the scene depth of the foggy image. After filtering, a simple, yet exact and effective, car detection algorithm is executed to confirm the presence or absence of a car in the processed image. The system is fairly robust and although all images were obtained from existing sources, the proposed algorithm is expected to perform equally well with any side-view image of a car in the presence of heavy fog and under real conditions