Traffic Light Detection Based On Adaptive Filter

Traffic detection play an important part in intelligent traffic system and vehicle.If we can do it efficiently,it is possible for a man with achromatopsiato safely drive on the road.If we want to apply it in our daily life,two important things need to be solved: 1)Real-time and high accuracy 2)Robust to complicate environment.So far,there are fewresearches that can satisfy those requirements.Most of the existing methods carried out in real-timebased on image processing method.But those methods rely on the quality of the image.Traditional CMOS camera cannot meet there requirements.So those methods are not robust tocomplicated environment.Traffic light detectioncontains two stages: 1)the candidate extraction stage 2)recognition stage.For candidate extractionstage,we propose a traffic light detector based on adaptive filter,it minimize the energy function of background,and introduce a equality constrain for the response of traffic light region,by introducing the relaxation variable and the least square norm regularization,we can solve it efficiently by interior method.And for recognition stage,we combine HOG descriptor and local RGB histogram,and design a cascade SVM classifier for traffic light semantic recognition.Because there is no a benchmark for us to train and test our recognition model,wecollect and label a lot of data by ourselves,and do lots of experiments on this datasets.We train our model on our datasets,and compared with Charette’s and Gong J’s methodon database of Robotics Centerof MinesAristech.The experimental result shows that,our method can achieve good recognition results,and is very robust toillumination variation and applicable on real road condition.