Detection of Traffic Panels in Night Scenes Using Cascade Object Detector

摘要

Detection and recognition of traffic panels (DRTP) in images is still a challenge in computer vision due to the huge amount of different conditions that are present on the road like cluttered scenes, light changes, occlusion, blurring and camera vibrations. The main purpose of this work is to contribute to DRTP since so far the papers reviewed resolve this topic for daylight scenes and it can not be guaranteed that those systems work for night scenes. This paper presents a computer vision system capable to identify traffic panels(TP) in night scenes. In preprocessing stage original image is resized up to 74% less, then is cropped to obtain the upper two thirds of the image and finally is converted into grayscale. In processing stage the output image from the last stage is processed by a cascade object detector (COD) to identify regions of interest (ROI). Each ROI is classified as TP or no TP in the classification stage using also a COD. In last stage each ROI classified as TP is highlighted with a bounding box. Two COD were trained, one using Histogram of Oriented Gradients (HOG) features and other using Haar features. The results of different combinations of such COD for preprocessing and processing stage are compared to obtain the most suitable combination to identify TP. An analysis of the results has determined that a combination of COD using Haar features and a COD using HOG features for preprocessing and processing stages respectively, is the best combination to detect TP.