Classification of Local Structures in Airborne Thermal Videos for Vehicle Detection

In this paper airborne thermal videos are used to detect vehicles. The movement of the camera is estimated from the optical flow using projective planar homographies as transformation model. A three level classification process is proposed: On the first level the eigenvalues of the squared averaged gradient are used to extract interest locations that can be put into correspondence with other such locations in subsequent frames of the video. These are subject to the second finer level of classification. Here we distinguish four classes: 1. Vehicles cues; 2. L-junctions and other proper fixed structure 3. T-junctions and other risky fixed structure. 4. A rejection class containing all other locations. This classification is based on local features in the single images namely Fourier coefficients. Only structures from the L-junctions class are traced as correspondences through subsequent frames. Based on these the global optical flow is estimated that is caused by the platform movement. The flow is restricted to planar projective homographies. This opens the way for the third classification. The vehicle class is refined using motion as feature. Inconsistency with the estimated flow is a strong evidence for movement in the scene.