🎓 This repository holds my masters dissertation project submitted in 2023 at the North-West University, South Africa titled 'Novel Data Augmentation Schemes for Pose Classification Using a Convolutional Neural Network'.
📓 Link to dissertation (to be added once link is available)
💽 Link to Dayta NWU Repositroy
Published research related to this project:
📘 SATNAC 2019 Heuristic Data Augmentation for Improved Human Activity Recognition
📘 SATNAC 2021 Colour-based encoding schemes for improved human pose recognition using a Convolutional neural network
Ensure that the following prerequisites are installed:
- NVIDIA GPU with CUDA support
- NVIDIA CUDA Toolkit
- cuDNN library
- Python 3.7.3
- Keras
The relevant Python libraries and dependencies can be installed using the requirements.txt
pip install -r requirements.txtThe first set of experiments act as a proof of concept to establish the viability of augmentating pose data with colour information to improve the capacity of a CNN to perform pose classification.
In these experiments, a curated image set was pre-processed by OpenPose to localise 18 body parts and joints across the captured human silhouettes. A baseline image set is generated by plotting the key points as white pixels onto a 32x32 image against a black backdrop. Six additional image sets are generated that incorporate colour, blending colours of overlapping key points, and tinting to reflect the OpenPose localisation confidence.
The second set of experiments expands on the findings of the preliminary experiments which demonstrated that colour acted as a salient feature in pose classification.
In these experiments, a video dataset designed for fall detection is pre-processed by OpenPose to localise the same 18 body parts and joints. The key point colour association is based on four different colour wheels which encode supplemental information differently based on the available spectrum of colours and their structural arrangement.
