Lagrange-based Video Encoder Optimisation to Enhance Motion Representation in the Compressed-Domain

Compressed-domain content analysis enables fast content-based video applications. However, conventional encoder implementation, limited to optimising data compression, does not necessarily result in content representative compressed features. In this paper, we present a novel method for video encoder optimisation that enhances reliability of content representation in the compressed video. We propose a configurable encoder model to jointly optimise bit-rate, distortion and content representation of the encoded video. In encoder optimisation for motion estimation, we analyse the behaviour of motion representation and compression efficiency over a range of content complexities. A mathematical model to control the extent of noise in selected motion vectors is presented. We demonstrate that the compressed domain motion information can be enhanced without incurring a rate-distortion overhead. The computational overheads are marginal and can be eliminated by exploiting the inherently parallel nature of the proposed optimisation model.