Motion ID 2.0 is a multimodal authentication system that combines gesture-based and voice-based identification using motion-captured data from 36 participants across various VR headsets. 

We implemented a gesture recognition pipeline using machine learning models such as Random Forest, XGBoost, and LightGBM, achieving a balanced accuracy of 92.75% and an F1 score of 91.88%. 

By incorporating Euclidean-based features, I further improved the performance by 3.11% in balanced accuracy and 3.69% in F1 score. The system achieved a 3% Equal Error Rate (EER) for user identification, which was further reduced to 1.5% with the inclusion of Euclidean features. 

To enhance security, I developed a speaker recognition system inspired by NVIDIA’s SpeakerNet (NeMo toolkit), based on the CAM++ approach, using a Dense Time-Delay Neural Network with channel attention to extract robust speaker embeddings. 

This project showcases my skills in machine learning, PyTorch, scikit-learn, shell scripting, and version control, along with my interest in AR/VR and biometric authentication.

This project focused on replacing expensive matrix multiplication with logical AND operations and Binarizing the input using standard datasets like MNIST. 

We implemented a K-means classifier and compared its results with feed forward Neural Networks (NN), CNNs, and their binarized versions. The results showed that BNNs achieved 96% accuracy while only occupying 1.7 MB of RAM, whereas standard NNs required 13 MB of RAM during training. 

This suggests that BNNs could be a more memory-efficient alternative to traditional neural networks.

This innovative project incorporated path planning, obstacle avoidance, navigation, and automatic speech recognition capabilities. To further enhance user control, the robot could be operated through a mobile application. 

We created a realistic simulated environment for testing the robot's functionalities using ROS, Gazebo and Blender. 

The project culminated in the authoring of a research paper published in the IJMERR November 2022 Issue, showcasing the project's design and functionalities.