Heuristic data augmentation for improved human activity recognition

This presentation was delivered at the SATNAC 2019 Conference, hosted at the Fairmont Zimbali Resort, Ballito, KwaZulu-Natal, South Africa. 1-4 September 2019.

[Abstract] Human pose estimation has been an important area of research in recent years due to its applicability in various everyday scenarios. Video surveillance is one such application where people detection and pose analysis can support safety and risk monitoring. Modern pose estimation implementations have attained notable success in terms of accuracy and efficiency thus affording the opportunity for viable solutions that enable such systems. In this study, data augmentation is employed to help improve real-time human activity recognition such as fall detection and other smart surveillance applications. Different augmentation techniques are applied to joint locations recovered from a pose estimator. The representation of the mapped joints is altered to include colour, the blending of colours where joints overlap, and the tinting of colours based on the degree of confidence for their approximated position. These augmentations and their combined effect are comparatively evaluated to yield the pose representation that is most beneficial for image classification when using a convolutional neural network. The greatest improvement in classification accuracy of 5% was attained using a combination of the examined augmentation techniques. These techniques can be diversely applied in different monitoring applications to enhance the detection of machine-learned behavioural patterns defined by pose estimations.

[Conference paper] Du Toit, J.S., Du Toit, J.V. & Kruger, H.A. 2019. Heuristic data augmentation for improved human activity recognition. In: Lewis, J. & Balmahoon, T., eds. Proceedings of the 2019 Southern Africa Telecommunication Networks and Applications Conference. SATNAC 2019, Ballito, South Africa. SATNAC. pp. 264-269.