Research On Parking And Abandoned Object Event Detection Algorithm Based On Traffic Video

It is necessary to detect the events of parking and abandoned object on the road real-timely and accurately,which is of great significance to avoid traffic accidents.Most existing detection algorithms not only need to construct the background model accurately,but also use 2D image features(such as area)to distinguish the target type.These algorithms have great dependencies on background,and the target types are not quite accurate.Therefore,this paper proposes a detection algorithm of parking and abandoned object,which uses the real 3D information of the target to distinguish the target type.Firstly,the preliminary detection of anomaly area is carried out based on state evolution.Secondly,it determines the parking and abandoned object area through two-way tracking of the initial abnormal area and uses the seed filling algorithm of eight-neighborhood to segment the parking and abandoned object.Finally,it uses three types of methods to distinguish parking and abandoned object.The first method is realized by using the difference between the projection velocities of the feature points in the region to determine the relative height between the feature points.Then,the parking and abandoned object can be distinguish by the different height.The second method is to distinguish the parking and abandoned object by calculating the suitability of 3D model between the projection area and the target area.The third method is to set up some different height of the reverse projection planes in the region,and the corresponding inverse projection map can be calculated.Then,it determines the target length,width and height through the different height of the inverse projection maps,and the parking and abandoned object can be distinguished by the information of long,wide and height.The algorithm has been tested in the scenes of tunnels,highways,urban expressways and rural roads.The test results show that the algorithm can detect the parking and abandoned object effectively,which has low miss and false detection rate.Moreover,it satisfied the real-time requirements.