Tracking small falling objects in video

This thesis identifies falling flowers and leaves in color video sequences of jacaranda trees. The spatial resolution is 320x240 pixels, and the temporal resolution is 15 frames per second. Tracking objects in video is a challenging computer science problem with many potential applications. For stationary cameras, the most common techniques utilize background subtraction for motion segmentation. Background subtraction attempts to identify the dynamic foreground by statistically modeling the static background. It acquires the foreground pixels by subtracting the background model from each incoming frame. In all but the most trivial examples, background models update the background with incoming frames, making the process adaptive. This paper uses a popular background subtraction technique called Mixture of Gaussians, which adaptively models each pixel with several (typically five to seven) Gaussian distributions. From the identified foreground pixels, regions are determined using the standard techniques of region labeling and morphological closing. The application then tracks the identified foreground regions between frames by relying on assumptions of the behavior of falling objects. The primary challenge faced in this paper is the small size of the objects, often consisting of only four or five pixels. It is often difficult to distinguish objects from noise inherent in the camera and the background model. The compiled results are promising, with a high rate of correct identification of events. I discuss reasons for false positives and negatives in detail, and provide several recommendations for future improvements based on these observations