Multi-Vehicle Detection And Tracking In Aerial Video

Multi-target detection and tracking has become a hot research field in computer vision in recent years, and the multiple target detection and tracking algorithms now has some shortcomings and deficiencies, such as different objects existing characteristics of its own, there is a complicated relationship between them, environment influencing the movement targets, object mutual occlusion, illumination changes, the noised image also can lead the missing and error detection to the break off of the tracking trajectory. The dense city environment is more likely to cause detection and tracking error. In order to more accurately describe the targets motion and predict the future objectives, this paper proposed a multiple target detection and tracking algorithm based on aerial video, the first step was video stabilization, then combined the three-frame subtraction and the background subtraction algorithm to get the foreground and background of the video, detecting and recognizing the moving objects, finally we extracted the difference feature attributes of the objects and used hyper-graph to match the objects between images to track the objects.Video jitter caused by outside environment, equipment and people will lead to a great deviation in multiple target detection, an algorithm of video stabilization based on optical flow was proposed. The algorithm used the optical flow to find matching points between frames and calculated affine formation matrix for video stabilization, and we can get an ideal effect. In practical application, the background subtraction algorithm uses GMM to get the background model but the background model may contain temporary stationary objects, the result is not ideal, and the three-frame subtraction algorithm exists the problem of "ghost" and "hole". This paper combined the advantages of the two algorithms, and we got a better result. After detection of the moving targets, we should extract the features of moving objects, such as color, speed, contour and texture, and calculate the difference between targets, such as distance, speed difference, then calculate correlation weight between goals between frames based on the hyper-graph matching, and matched the objects in the fames before and after according to the weight. The experimental results show that the algorithm not only worked in real-time detection and tracking while the accuracy is improved obviously, but also has strong robustness.