| Machine vision technology is a hot research topic in recent years, and intelligent object detection and tracking have a very important position. However, most of the target detection algorithms in dynamic background are not accurate and real-time low. When the target is affected by occlusion, rotation, illumination and other more factors, general tracking algorithms are also prone to loss of tracking, dithering and so on. And limited by the algorithm’s high complexity, producing large amount of computation, detection and tracking algorithms are difficult to escape from the large-scale computer system, so that it is hard to be applied in a variety of mobile devices.Mainly aiming at these problems, this paper puts forward a global vector estimation algorithm based on motion cancellation of features matching, to compensate the movement background, for the accuracy of the detection of moving targets. Secondly, combined with Kalman filter, the method proposed in order to track is based on the local characteristics updating of Bhattacharyya coefficients estimation. Experiments prove that it is good robustness and real-time. Finally, this algorithm is used to develop an application with Qt cross-platform development tool, which is successfully transplanted to embedded operating system of Davinci DM6446 development board. The main work of this paper include the following aspects:1. Basic theory and new algorithms of detection and tracking in recent years are learned and researched on, and classical popular feature point operators are analyzed in detail. And tracking methods based on Kalman filter and particle filter are introduced.2. A target detection algorithm in dynamic scene is proposed. Although the using of classical algorithms, such as SURF and SIFT, to extract the feature points, have a certain degree of stability for changes of scale, rotation, brightness and many other factors. But more of these are not conductive to real-time detection due to the slow generation of feature points. Assumed that target detection in continuous video sequences is relatively stable, the Freak operator based on FAST in this paper is used to select feature points to improve detection rate. For the situation that when the target shape is larger, the feature points generated by local affect estimation of global vector leading to failure detection, RANSAC algorithm is used to eliminated the error matching, combined with grid method and motion cancellation to obtain background after compensation. Finally, multi frames difference, morphological filtering, threshold segmentation and more methods are able to be used to accurately detect and extract the target area.3. A target tracking algorithm in dynamic scene is proposed. In video sequences the objects are usually strong maneuverable, so Freak feature operator is selected for the better matching and effectives in this paper. For the problems of occlusion and jitter when tracking targets, Kalman filter is used to continuously estimate the position of targets. For the characteristic change caused by target deformation, the similarity of objects is estimated by Bhattacharyya coefficient in order to update local feature points. Experiments prove that the proposed algorithm in this paper makes accuracy and real-time of tracking remarkably improved.4. A embedded software of tracking system is designed and developed. The embedded development environment is constructed by DM6446 development board and kit provided by ruvia corporation. Finally, with QT embedded development tools in the Linux operating system environment, the algorithm proposed in this paper is developed into a visual software successfully transplanted to embedded device. |
PDF Full Text Request