Cross-domain traffic scene understanding by motion model transfer

Abstract

This paper proposes a novel framework for cross-domain traffic scene understanding. Existing learning-based outdoor wide-area scene interpretation models suffer from requiring long term data collection in order to acquire statistically sufficient model training samples for every new scene. This makes installation costly, prevents models from being easily relocated, and from being used in UAVs with continuously changing scenes. In contrast, our method adopts a geometrical matching approach to relate motion models learned from a database of source scenes (source domains) with a handful sparsely observed data in a new target scene (target domain). This framework is capable of online ''sparse-shot'' anomaly detection and motion event classification in the unseen target domain, without the need for extensive data collection, labelling and offline model training for each new target domain. That is, trained models in different source domains can be deployed to a new target domain with only a few unlabelled observations and without any training in the new target domain. Crucially, to provide cross-domain interpretation without risk of dramatic negative transfer, we introduce and formulate a scene association criterion to quantify transferability of motion models from one scene to another. Extensive experiments show the effectiveness of the proposed framework for cross-domain motion event classification, anomaly detection and scene association.