Colour-based encoding schemes for improved human pose recognition using a Convolutional neural network
This presentation was delivered at the SATNAC 2021 Conference, hosted at the Champagne Sports Resort, Central Drakensberg, KwaZulu-Natal, South Africa. 21-23 November 2021.
[Abstract] The goal of computer vision is to provide computers with the perceptual capability to process and understand visual data and is achieved by synthesizing information from raw videos and images. Humans are among the most frequently analysed subjects of computer vision because activity recognition and pose identification are applicable in various critical industries. This study examines a novel data augmentation technique that can aid in the machine-learned interpretation of the human form for improved pose recognition by superimposing joint markers on humans detected in video footage. The joint locations are derived from a pose estimator and are supplemented with additional information through the use of colour. The keypoint colour augmentations are applied based on either a radial or ringed colour wheel, which is notionally intended to encode spatial information based on the position of the joint. An improvement in classification accuracy of up to 11 percentage points over a baseline model was achieved when applied to a pose dataset and classified using a convolutional neural network. Furthermore, different augmentation schemes were found to favour the recognition of certain poses over others. These augmentation schemes can be diversely applied in human activity recognition and tailored to foster improved accuracy for pose-dependent classification tasks.
[Conference paper] Du Toit, J.S., Du Toit, J.V. & Kruger, H.A. 2021b. Colour-based encoding schemes for improved human pose recognition using a convolutional neural network. In: Smuts, M. & Moorcroft, R., eds. Proceedings of the 2021 Southern Africa Telecommunication Networks and Applications Conference. SATNAC 2021, Central Drakensberg, South Africa. SATNAC. pp. 178-183.