Automated Learning Video Analysis (ALVA)

We propose to develop a system that will automatically track targets based on characteristic velocity differentials between the target and the background. We leverage our existing system for feature detection and tracking and show that it can produce recognizable discrepancies in the kinematics between a moving background and a moving target. We will augment the discriminating power of this kinematic information with visual cues extracted from the video around the tracked centroid. The resultant feature vector will be used both for finalizing the tracking and for identifying events from the video sequence. We present sample data which suggests that the kinematic information alone might be sufficient to automatically recognize the object and background from the image without additional input required from the engineer. We demonstrate an intriguing transformation in which the video data cube can be projected to substantially simpler 2D images that are amenable to further image processing. We discuss how the reduced and well formulated information stream can be processed through the use of pattern recognition and Neural Networks to recognize and segment discernable events. This report demonstrates the potential of our approach on both ground based, flight chase and cockpit video sequences.