Traffic Incident Detection Algorithm Based On Machine Vision

Traffic accident detection is one of the most important issues for intelligent transportation systems.Real-time and robust detection methods of traffic accident can make a great contribution to minimize the casualties and property losses.With the rapid development of intelligent transportation systems,traffic accident detection employing computer vision and image processing has attracted much attention recently,and significant progress has been made by researchers in this field.However,due to the complex traffic environment,these proposed methods still have rather limited capability for practical application.To enable the rapid and accurate detection of traffic accident and meet the need of practical application,there are a number of challenges which need to be addressed.Such a detection method should:(i)deal with complex transportation environment,which includes bad weather,different illumination conditions,complex road conditions and traffic participants,(ii)deal with different traffic flow states,which mainly reflect to the level of the traffic congestion,(iii)operate in real-time.This paper investigates real-time and robust traffic accident detection in a traffic video stream.Tradition methods are either not fast enough or unstable under complex working environments.We propose a novel video-based traffic accident detection method.The key observation of the method is that the traffic accident brings local vehicle confusion on motion directions at the same time it causes the disorder of the global traffic flow.This method includes three steps.First,the optical flow field for each video frame is constructed.Then,the global traffic flow descriptor and the local orientation map are constructed based on the optical flow field,for the global and local features of traffic accident,two Gaussian models are established to detect the traffic accident in the video frame and locate it.Finally,an inspection module is built to further verify the occurrence of the accident and exclude the possibility of false alarms.The advantages of our method include the real-time operation,high accuracy,low false negative and low false positive,and the robustness to the different environments and illuminations.We carry out an extensive quantitative evaluation which confirms the above advantages and shows a significant improvement over state-of-art approaches.