Implementation of motion detectors: A case study

Two types of motion detectors, a Reichardt/correlation-based detector and an energy filtering-based detector, were implemented in order to evaluate their applicability in a real-time computer vision scenario. The detectors were evaluated both on controlled artificial stimuli and, more importantly, on real-world video sequences of moving faces. The Reichardt correlation model requires the least amount of computational effort, but its output is sensitive to noise, especially in bright parts of the image. To address these issues, two variants of this model were implemented. In the first variant, an average of local pixel values was used as input to the detector in order to increase robustness. In the second variant, the normalized cross-correlation (NCC) between image regions was used as input to the detector, which increases computational effort but makes the output contrast invariant. We also implemented a motion detector based on the Adelson and Bergen’s oriented spatio-temporal energy filtering model. This model is contrast-sensitive, but provides more stable and accurate output, at the cost of greater computational complexity. The contrast sensitivity of detectors was found to be an advantage when a high-contrast area is of interest, providing a certain degree of built-in segmentation. We propose an architecture for motion analysis consisting of a first stage of coarse motion detection using the more computationally-efficient averaged Reichardt detector, followed by a finer analysis of specific regions of interest using the NCC Reichardt or the energy-based detectors