Computation and performance trade-offs in motion estimation algorithms

Real-time video/visual communication applications require trade-offs in terms of processing speed, visual image quality and power consumption. Motion estimation is one of the tasks in video coding that requires significant amount of computation. Block matching motion estimation algorithms such as the three-step search and the diamond search algorithms are being used in video coding schemes as alternatives to full search algorithms. Fast motion estimation algorithms reduce the computational complexity, at the expense of reduced performance. Special purpose fast processors can be employed as an alternative to meet the computational demand. However, the processing speed comes at the expense of higher power consumption. This paper investigates motion estimation algorithms and presents the computational, and performance trade-offs involved in choosing a motion estimation algorithm for video coding applications. Fast motion estimation algorithms often assume monotonic error surface in order to speed up the algorithm. The argument against this assumption is that the search might be trapped in local minima and may result in a noisy motion field. Prediction methods have been suggested in the literature as a solution to avoid these local minima and noisy motion field. The paper also investigates the effects of the monotonic error surface assumption as well as the appropriate choice of initial motion vectors that results in better performance of the motion estimation algorithms