Model-less and model-based computationally efficient motion estimation for video compression in transportation applications

Block-based motion estimation is an important component in many video coding standards that aims at removing temporal redundancy between neighboring frames. Traditional methods for block-based motion estimation such as the Exhaustive Block Matching Algorithm (EBMA) are capable of achieving good matching performance but are computationally expensive. Alternatives to EBMA have been proposed to reduce the amount of search points by trading off matching optimality with computational resources. Although they exploit shared local spatial attributes around the target block, they fail to take advantage of the characteristics of the video sequences acquired with stationary cameras used in transportation and surveillance applications, where motion patterns are largely regularized; often, they also fail to yield semantically meaningful motion vector fields. In this paper, we propose two alternative approaches to improve the efficiency of motion estimation in video compression: (i) a highly efficient model-less approach that estimates the direction and magnitude of motion of objects in the scene and predicts the optimal search direction/neighborhood location for motion vectors; and (ii) a model-based approach that learns the dominant spatiotemporal characteristics of the motion patterns captured in the video via statistical models and enables reduced searches according to the constructed models. We demonstrate via experimental validation that the proposed methods attain computational savings, achieve improved reconstruction error and prediction capabilities for a given search neighborhood size, and yield more semantically meaningful motion vector fields when coupled with traditional motion estimation algorithms.