Research On Motion Segmentation Algorithm Based On Event Camera

Due to the limitation of the sensor's performance,the traditional camera can't acquire theoriginal image with good quality in the fast and low exposure environment,so the performance of traditional vision algorithm will be seriously affected in the special environment.As a new type of sensor,event camera has the characteristics of fast response,low delay and high dynamic range.However,the main datas of event-based camera recording are "event points",including pixel brightness change,location and time,which are difficult to be compatible with traditional visual algorithm.When the camera does not move,the event camera only focuses on the moving object.If display the events during a short time at their location,the image usually shows as the edge of the moving object.However,if camera moves,the scene information will be recorded together,so it is difficult to distinguish the motion foreground and background information.This thesis proposes a motion segmentation algorithm based on event camera using its low light perception characteristics,which consists of three parts.The first part is the model adaptive selection module,which determines whether the recorded traditional visual image can provide reliable information by comparing the difference between the event information and traditional visual image according to their edge.The second part is the optical flow prediction module,which uses APS domain optical flow prediction model based on traditional visual image and event domain optical flow prediction model based on event information which are respectively used to provide optical flow information such as approximate motion direction and speed of the scene and moving object in the case of good or bad illumination.The third part is the probability fusion module,which gives each event point a initial probability,and uses the previous obtained optical flow information to distort the location of event points,then adjusts the optical flow according to the contrast principle,and fuses the probability of event points in the process,which is used to distinguish foreground and background event points.In the third part,the author also design a post-processing algorithm to remove the noise and background information in the foreground events to get the final motion segmentation results.In the experimental part,a detection method is designed to determine the bounding box of the foreground object obtained from the motion segmentation results.The results show that,in the case of better illumination,the mean intersection-over-union of the moving object's bounding box obtained by the algorithm can be higher than that of the conventional tracking algorithm,and in the case of poor illumination if it's unable to obtain the reliable traditional visual image,only using event information,the algorithm can still get reliable motion segmentation results.