|导师||杨一平 ; 常红星|
|关键词||运动目标检测 航拍图像序列 高分辨率 无人机 精确检测|
The fast and accurate moving object detection under the unmanned aerial vehicles (UAV) platform is the key to the intelligence of UAV. It plays a very important role, no matter in target reconnaissance, fire attack and other military applications or intelligent transportation, police security and other civilian areas. The previous works on moving object detection from aerial videos are mainly designed for low-resolution videos and there are still many aspects need to be improved, such as the real-time capability, the adaptability to environment and the detection accuracy. Recently, to achieve a wide-area of surveillance with fine granularity, most of the UAVs prefer capturing high-resolution videos but can only process a few frames of the videos due to constraint from computing resource, making the existing video-based algorithms (need many continuous frames for detection) unsuitable for the high-resolution scenario. To solve these problems, this paper carries out systematic research on moving object detection from aerial videos, especially from the high-resolution aerial videos. Our main works include the following three parts:
(1) Considering that most of the existing background modeling based moving object detection algorithms suffer from the accumulation of the stabilization errors and parallax interferences, we propose a dynamic online background modeling based algorithm. By making stabilization between the current frame and the dynamic background model, we can reduce the accumulation of the stabilization errors. Considering that the dynamic background image is fuzzy and does not contain the information of moving targets, we propose to use a tracking based matching strategy to stabilize the background image and the current frame fast and accurately. In considering the frequent shaking of the UAV and the wandering detection situation, we build a background model which can record the information of the pixels near the frame boundary and better handle the objects and background near the fringe field of view. Experiments show that the proposed method can detect the moving objects more accurately with a high computational efficiency. Besides, it has a high adaptability to environments.
(2) In view of the existing moving object detection algorithms under the UAV platform are all designed for low-resolution videos, which often require many frames and are difficult to detect the moving targets in real-time, cannot tackle the motion detection task for high-resolution situation, we propose a feature classification based method for online moving object detection from high-resolution aerial image sequences. The proposed method detects the moving objects by fully utilizing the motion information of the feature pairs between two frames. By performing the detection in the feature pair domain instead of in the pixel domain, the proposed method has a very high computational efficiency. We design a whole detection system, which includes feature extraction and refinement, feature classification and consistency based clustering. We propose a tracking-based block extraction strategy and a feature matching pairing algorithm base on consistency and block matching, and design valid features for distinguishing the pairs on the moving objects from background pairs and design a consistency criteria for clustering. Experiments are made to validate the effectiveness of the proposed method. Experimental results show that the proposed algorithm can detect and locate the moving targets fast and accurately. The proposed algorithm has a very high computational efficiency and can leave valuable onboard computing resources for high-level vision tasks.
(3) Taking into account the common issues of detecting only part of the moving target, detecting one target as multiple targets and detecting multiple targets as one target for moving object detection algorithms, we propose an incremental feature organization and self-directed segmentation based method for moving object detection from high-resolution aerial image sequences. By the incremental feature organization based expansion and merging algorithm, time domain information is fully used to tackle the problems of part detection and detecting one target as multiple targets. By the GrabCut based self-directed segmentation algorithm, the spatial domain information is amply utilized to handle the problems of detecting multiple targets as one target and part detection. Experiments show that, the proposed method can better solve the problems of part detection, detecting one target as multiple targets and detecting multiple targets as one. With high computational efficiency, the proposed algorithm can greatly improve the performance of completely and accurately extracting the moving objects.
The works of this paper cover moving object detection from low-resolution and high-resolution aerial videos and they constitute a complete system for moving object detection under UAV platform. To our best knowledge, these works tackle the moving object detection form high-resolution aerial image sequences for the first time, which lay a solid foundation for embedding the moving object detection system into the hardware platform for practical applications.
|作者单位||Institute of Automation, Chinese Academy of Sciences|
|兰晓松. 高分辨率航拍图像序列中的运动目标检测方法研究[D]. 北京. 中国科学院大学,2017.|