Multiple moving object estimation in image sequences of a natural scene

Texture- and shape-preserving, velocity-selective moving object estimation in time-sequential imageries are investigated. A 3-D spatiotemporal filter is usually effective for velocity-selective processing of nonmaneuvering moving objects. However, its performance is limited by the inherent aliasing problem due to discrete space and time sampling of continuous signals of moving objects with various motion vectors. To complement the imaging effects due to spectral aliasing apparent in the spectral domain processing, the time-recursive temporal low-pass filter, which is based on the Kalman theory, is incorporated in parallel to the proposed 3-D spatiotemporal filter banks. This temporal low-pass filter is effective in adaptively separating relatively stationary backgrounds from all other moving objects. In other words, the false moving objects due to the imaging effect can be successfully suppressed using the binary masks of all moving objects that are obtained through the simple type of time-recursive temporal low-pass filtering. From the simulation using several real IR image sequences, not only for the multiple moving objects with various shapes and speeds but also for the noisy image sequences, the highly accurate texture- and shape-preserving, velocity-selective moving object estimation results are observed with a graceful degradation in accordance to the increased amount of noise.