Research On Video Based Traffic Lights Recognition And Shadow Eliminate And Its Application

Object recognition is an important research area in computer vision, while traffic lights recognition and shadow recognition are current central issue in object recognition. Real-time traffic lights recognition is widely used in driver assistant systems and Unmanned Ground Vehicles (UGV), while shadow eliminate is related to the effect of object tracking,which is the basis of many high lay algorithms of computer vision,such as classify and behavior description and image understand. Because of the complex of the background caused by motion of UGV and demanding real-time and traffic light blob being small in image,traffic lights recognition is a challenge work.The shadow eliminate is complicated due to the feature of shadow. Firstly, shadow is different from the background just like the foreground object, secondly, shadow and the object which the shadow belong to have the same movement when the object is moving.The shadow sometimes sticks with the object which the shadow belong to,sometimes separates from the object,which makes the problem more complicated.According to deficiency of current traffic light recognition algorithm in accuracy or in real-time, we propose an algorithm which is based on Lab color space and template match.This paper makes two improvements in real-time traffic lights recognition:method of modeling the color of traffic light and method of confirming the candidate area.In color modeling, the current algorithms model the color in RGB space or in HSV space, which need several thresholds.We observe that, the traffic lights only have three colors,so we just need three thresholds while modeling the color in Lab space,which is more proper in describe red,green and yellow, thus we promote the robustness of our algorithm.In candidate area confirm,the current algorithms,which only consider the candidate area's shape or use complex template, have low correctness of recognition or high overhead.We observe that,the traffic lights are wrapped by a black rectangle,so we design three traffic lights template to confirm the candidate area, which promote the recognition correctness while keeping real-time.According the deficiency of current shadow eliminate algorithm in robustness we propose an algorithm which is based on shadow characteristic.The shadow eliminate methods assume that an area cast into shadow often results in a significant change in intensity without much change in chromaticity. They set up several thresholds to find the shadow area, while the effect depends on thresholds.The methods lack robustness. We analyze the histogram change of the area after shadow cast,which show that area is gauss distributed,so we set up three Gauss Models to describe shadow. We compute the match degree of our model and the foreground area to confirm shadow. The algorithm is more robust because it needs fewer thresholds.In order to evaluate our shadow eliminate algorithm,we recode two out-door sequences,and download two in-door sequences from university of California, San Diego.The evaluate result show that,our algorithm eliminate the shadow effectively. Compared with other shadow eliminate algorithms,our algorithm gets a better effect by sacrificing a little overhead.In order to evaluate our algorithms,we run them on Hunan University's UGV. Our traffic lights recognition algorithm is part of the Hunan University's UGV, which is the champion of the first china "UGV future challenge" completion. In the completion our algorithm recognizes the traffic lights successfully. In order to evaluate our traffic lights recognition algorithm,we record three sequences from an on-vehicle camera.We compare the result with other algorithms.The experiment shows that,our algorithm recognize the traffic result in real-time.Compared with other traffic light recognition algorithms,our algorithm gets a balance in speed and effect.