Study On Multi-Object Detection And Tracking In Video Captured From Low-Altitude Platform

With the development of machine vision, more and more researchers of the intelligent transportation field increasingly concerned video traffic monitoring system on low-altitude platform. The core of it is using camera installed in the low-altitude airships or unmanned platforms to get traffic scene, detecting the traffic objects such as vehicles through video image processing technology and tracking the traffic objects to estimate their moving status, to realize the intelligent road traffic management. Comparing the general static video surveillance systems, it is a dynamic monitoring system. The mobility of detection platform, real-time change of scene, low resolution and mobility of the detection object, multiple object tracking and real-time application make the vehicle detection and tracking on low-altitude platform become a technical problem.In this paper, the following research is carried out against the difficulty of video traffic monitoring system on low-altitude platform.First, according to the mobility and real-time changes of the detection platform, a new technology on video stabilization combined with optical flow vectors is proposed. This method uses the image stabilization technology, which establishes the correspondence between the reference frame and target frame through feature matching, to make up the detection platform moving and the vehicle detection of dynamic background is turned into the problems of static background. Then a low computational complexity of the optical flow vectors is used for vehicle detection. The four algorithms are analyzed on videos captured by cameras on low-altitude platform. Gaussian mixture model approach on the assumption of static background, optical flow method on hypothesis of dynamic background and KLT tracking method on Tomasi are not well in detecting the moving vehicles, while the proposed video-based stabilization combined optical flow vectors method can better detect moving vehicles.Second, for the dynamic of the detecting objects and real-time requirements, a video stabilization on C_SURF feature extraction is proposed. The main step of C_SURF feature extraction is the same as SURF, except the main direction allocation and feature description take into account the color factor. Thereby the number of feature is reduced. The performance of SIFT, SURF, CSIFT and C_SURF is analyzed under four situations including image scale changes, image blur change, image view angle changes and image brightness changes. The best performance of the algorithms is SIFT, but the best efficiency is C_SURF. And the performance analysis of four algorithms on video-based stabilization algorithm shows that C_SURF efficiency is optimal when the video stabilization accuracy is considered basically the same.Third, aiming the low resolution for detection of objects, a new particle filters on color and shape joint characteristic is proposed to tracking multi-targets. According to the obtained object and assuming the target is uniform linear motion, a movement model is established on uniform motion and an observation model is established on color and shape joint characteristic. The proposed algorithm can effectively track multiple moving vehicles on the videos captured by cameras on low-altitude platform.