Research On Multiple Object Speed Estimation And Tracking Based On Object Detection And Optical Flow Calculation

In recent years,thanks to the flourish of deep learning on computer vision field,some traditional difficult tasks like object detection and optical flow calculation are provided with new solutions.However,these methods are in specific niche and mostly can not be used singly in complicated scenes,while synthesizing them could find out useful information to solve the comprehensive problems.About the dynamic characteristic field in computer vision,simultaneously achieving multiple object speed estimation and track tracking can obtain the historical and current state of motion of multiple objects,which are the key information in application like road accident analysis assistant,athlete tactics pitch execution analysis and so on.Currently,there is no common technology in solving multiple object speed estimation based on computer vision,while multiple object tracking has been developed in recent years and is still in exploring stage in computer vision field.Based on above background,this dissertation aims to combine two types of CNN: object detection one and optical flow calculation one,to simultaneously achieve multiple object speed estimation and multiple object tracking.As for multiple object tracking,this dissertation designs a multiple object tacking algorithm based on object detection and optical flow calculation(ODFMT).It uses optical flow to filter candidate boxes,uses optical flow and Kalman Filter to achieve track prediction,uses optical flow contour and other basic image features to achieve feature matching.The experiment result shows the algorithm has advantage in running speed and track prediction compared to the current mainstream algorithms.As for speed estimation,this dissertation designs a speed estimation based on object detection and optical flow calculation(ODFSE).It uses object location together with cluster method to calculate out object optical flow,then converts it to speed by means of optics theory,camera calibration and priori knowledge.The experiment result shows the algorithm is able to estimate object speed in condition of camera moving,whose accuracy is acceptable in most application field s.Based on the two algorithms above,this dissertation designs and implements a multiple object speed estimation and tracking system.It uses the image of a single camera as data resource and simultaneously outputs the estimated speed and track of multiple objects.