Moving Target Trajectory Reconstruction Based On Stereo Vision

Detection and tracking of moving targets has been a hot spot and one of the difficulties in research of computer vision, in intelligent transportation, robotics,human computer interaction, weapon guidance, industrial control, medical treatment,etc, had been used more and more. Its purpose is to capture interested targets in video, provide target location and track information for subsequent processing.This paper discusses the 3-d reconstruction of moving target trajectory based on stereo vision. First we use the improved Vi BE target detection algorithm to build the background model based on image color and depth information, according to the image color and depth changing patterns detect the moving object in video, extract the precise target image. Then we use the improved KCF method tracking moving target, while maintaining the adaptive capacity of original KCF algorithm, aiming at the condition of long term occlusion of the target to make the improvement, make the tracking process more robust, avoid the the tracking drift problem. Finally we based on stereo vision principle to calculate the 3-d target position, then synthesis the target trajectory with the calculated trajectory points.In moving target detection stage, the traditional Vi BE(Visual Background Extractor) moving target detection algorithm often produce some unwanted areas,such as Ghost area, target shadow and illumination changes, etc.This paper proposes an improved Vi BE detection algorithm to identify foreground and background base on the relative changing patterns in depth. Firstly, image color and depth information is used to establish the double background models;Then using color background model to detect possible foreground areas in the current frame; Finally,depth background model is introduced to identify false foreground area, and gives the updating methods of background region, ghost area, shadow area. This method is simple, with small amount of calculation, can achieve a higher recognition rate of the false target area, and can handle several issues simultaneously, also has a mechanism to deal with the invalid depth data and the depth noise, which improved the reliability. Experimental results show that the algorithm can effectively eliminate ghost area and target shadow, illumination change areas etc. Compared to Vi BE algorithm, the average precision is increased by 34.68%, in dealing with the resolution of 640*480 pixel color images, the average frame rate is 34.96, can fully meet the requirements of real-time processing.In moving target tracking stage, the traditional KCF(Kernelized Correlation Filter) target tracking algorithm would drift in dealing with long time occlusion. In order to solve this problem, this paper puts forward the use of the depth distributionin target area to identify the target occlusion, and the use of dynamic adjustment of model update rate and synthetic training set to reduce negative samples. First of all,based on the depth distribution of the target area calculate the target occlusion ratio;Then use the occlusion ratio to adjust model update rate, lower the model update rate when occlusion happens; Finally, we compose the target training set using the occlusion mask when target occluded, to avoid drift problem. The improved KCF algorithm reduces the negative sample noise effectively, keeps the stability of the tracking, is able to deal with the long time occlusion.Finally we introduced the clear or blurred target trajectory reconstruction method. For clear target, the use of image automatic registration technology based on target feature points detection and area matching, can accurately reconstruct the points on the target. The combination of the depth information with the target tracking algorithm, improves the integrity and accuracy of the target tracking and the trajectory reconstruction, and then through the calculation of the points on the target, the target space position is obtained. For motion blur scenario, using direction of differential method to calculate the instantaneous target motion direction, Hermite interpolation method is used with the target location and the directional derivative information to interpolate the centroid trajectory. Several points would be produced by the centroid trajectory matching procedure of each frame. In order to enhance the accuracy of trajectory reconstruction at the end, to reduce the random errors in the measurement process, we discuss the feasibility of increasing the number of cameras at different angles. At the end, we use the nonlinear curve fitting technology to produce the moving target trajectory based on the target position obtained at different time.