Moving Objects Detection And Visual Navigation For Aerial Robots

The research of aerial robots has a lot of potential applications and it became one of the most active research areas in robotics community. The research of visual perception and navigation of aerial robots are two major issues in this area, and they are very challenging.Aerial robots are dynamical platforms and they move in3D scenes, therefore the research of computer vision of aerial robots has the following challengings:(1) Different from ground robots, aerial robots could move freely in3D environments, so the images captured by the cameras on them are more complicated (for instance, the altitudes of aerial robots could be changed dramatically in the experiments and this is different from ground robots), thus a lot of visual perception and navigation methods could not be incorporated perfectly in aerial robot platforms.(2) Different from ground robots, aerial robots are unstable platforms. Therefore they usually move rapidly and the control frequency would be very high, thus they demand high speed image collection and processing techniques. Further more, there would be much more noise in these images, therefore aggregates the difficulty of visual perception and navigation.(3) Different from traditional robots, the pose estimation calculated by the computer vision methods is usually integrated in the attitude control of the aerial robots, and this demands the robustness of the pose estimation methods.Considered the challengings above, and based on the platform of aerial robots, this thesis mainly focuses on the research of moving object detection and visual navigation in dynamic scenes (for dynamic scenes, it means that there exist some other moving objects in the camera scenes other than the background.):(1) It has proposed a new motion segmentation based moving object detection method in dynamic scenes. This method employs an effective mechanism to take advantage of the results of3-D motion segmentation in order to track the background, thus segments the moving objects. Experimental results show that this method could overcome the weaknesses of the state-of-the-art methods and detect moving objects from dynamic scenes.(2) It has proposed a new appearance based method for moving object detection in dynamic scenes. This method takes advantage of the "key-frame" segmentations, tracks them through the other frames, builds the appearance models and gets the optimized segmentation using graph-based optimization. Experimental results show that this method could be applied to complex scenes and detect the moving object accurately. Since this method does not need to extract key-points from the frames, it could detect small moving objects effectively.(3) It has proposed a new camera pose estimation method in dynamic scenes. This method extracts key-points from the frames, matches them into trajectories, employs3-D motion segmentation method and background tracking to extract background points, and then uses multiple view geometry method to estimate the camera pose. Experimental results show that this method could estimate the camera pose accurately in dynamic scenes, however, traditional method would induce too much error for the estimation.(4) It has proposed a new map stitching method in dynamic scenes. This method undistorts the input images, removes the moving object from the background using the method above, extracts and matches key-points, and estimates the transformations and stitches the images together. Experimental results show that, traditional method could also get stitched map, however, the map includes parts of the moving objects; in contrast, the proposed method could effectively remove the moving objects and obtain a more accurate and reasonable stitched map.