Nighttime Vehicle Detection With Adaptive ROI Using Binary Samples

Since 21th century,intelligent driving technology for vehicles has developed rapidly.At present,self-driving cars,such as Google,Tesla,Waymo,General Motors and Uber,have been the pioneer of new traffic facilities.Self-piloting automobile has become a new trend.Vehicle detection is the foundation of autonomous vehicles or Advanced Driver Assistant System(ADAS).Currently,it is mainly based on passive sensors such as Radar and Lidar,which can meet the requirements of high accuracy and real-time.The commercial semi-automatic driving systems,such as the Tesla Autopilot,Cadillac Super,Mercedes Benz Drive Cruise Pilot,have an average price of $100 thousand to $60 thousand.The high cost hinders popularity of the automatic driving systems in civilian use.In addition,sonar and millimeter wave radars can detect the surrounding targets quickly,but cannot identify distant or small targets.When the traffic is congested,the interference between device signals and the cumulative accumulation of errors affect operation of the systems.At present research,vehicle vision technology based on computer vision uses inexpensive cameras to collect high-resolution traffic images.As an alternative,it can greatly reduce the cost of automatic driving system and provide abundant visual information for target recognition.Most of the current researches on vehicle detection based on vision focus on the daytime mode,while the nighttime urban road lighting is complex and uneven illumination,which challenges the vision-based nighttime vehicle detection method.Vehicles(lamps)show irregular bright clumps in the traffic image.The shapes,colors,edges,textures and corners of the vehicles in front,which are used for daytime vehicle detection,are missing in the low-illuminating nighttime traffic images.At the same time,the regular motion characteristics of vehicles(lamps)cannot be extracted.Therefore,most vision-based daytime vehicle detection methods are not applicable to the nighttime mode.In the process of driving at night,the lack of light results in limited vision.Drivers cannot predict the driving trend of the front vehicles accurately and quickly.It is easy to cause operational errors due to subjective factors such as fatigue and distraction,so that there is great demand for Intelligent Driving at night.To solve the above problems,this paper presents an approach using binary spotlight Haar-like features to train the cascaded AdaBoost classifier for nighttime vehicle detection.The detection result is further compensated by a spotlight maximum similarity criterion between successive frames.Aiming at the complex lighting problems of nighttime urban road,we also propose to define a region of interest(ROI)by detecting the line segments where the streetlights locate.An improved binarization scheme is adopted to further remove the light interference for image preprocessing.Experiments show that our method has an average recall rate of 93.78%;the use of binary samples reduces 35%training time;adaptive ROI removes about 75%of the background detection area.We demonstrate that our algorithm has a real-time performance and high accuracy,and it also offers convenient adaption for vehicle detection to various types of disturbances on urban road during night.The experimental results show that the proposed threshold processing method and the ROI determination method are suitable for the pre-processing of nighttime traffic images,and the vehicle detection method based on binary sample classifier has good real-time performance and recall rate,with adapting to various road conditions for the night vehicle detection.