Tracking Moving Objects from a Moving, Autonomous Helicopter

In this paper we address the problem of tracking motion in diverse video when the camera may also be in motion. Our implementation handles this diversity by integrating a variety of approaches and selecting the appropriate approach or approaches for the input material using an intelligent algorithm on a frame‐by‐ frame basis. Accuracy is over 75% for some inputs and can handle up to 10 frames per second. 1 Introduction Many factors affect the detection and tracking of motion in video. Video capture devices differ in terms of calibration, aspect ratio, resolution, and numerous other factors. Also the capture device could be mounted in different orientations relative to its field of view, could change orientation during the capture process, and/or could move with its platform. The object tracked could be a small or large component of the frame size. It may not have features that can be tracked or it may have comparatively fewer features than its surroundings. Its motion might be constant or could vary with acceleration or present with a non‐ linear path. There might be multiple objects to track. While there has been a reasonable amount of work done on motion‐tracking (see works cited), the problem still remains open and research on the subject is ongoing. Our projected approach, one which combines traditional motion compensation techniques and object tracking techniques with probabilistic models (i.e. particle filters), is a current area of interest. We hope our implementation will yield interesting results. This paper presents our process and results thus far. Our sparse optical flow algorithm consists of feature detection, feature tracking, camera motion compensation, and an intelligent tracking algorithm. While this system achieves good results, we have also implemented an algorithm utilizing dense optical flow methods for comparison and experimentation. Where appropriate, we provide results from both implementations, and realize that our final algorithm may incorporate aspects of both. In this paper we discuss these approaches and provide preliminary qualitative and quantitative results. Section 2 describes previous work. Section 3 presents our approach, formally describing its components and their interaction. Section 4 offers qualitative (video with tracking information superimposed) and quantitative (accuracy and fps metrics) results. Section 5 discusses how the project will progress in the last half of the quarter. Section 6 concludes. 2 Background The principal problem in tracking motion from moving platforms consists of two sub‐