Team:- ADAS
★ Developed a novel HM-LSTM model to predict early drowsiness using blink features with 92% accuracy, optimized using TFlite, and deployed it using ArmNN on low-cost ARM hardware.
★ Improved speed of object detection models by 2.5X using quantization, pruning, and TensorRT without any accuracy loss.
★ Created a MobileNetV3 backboned multi-branch network to detect vehicles and segment lanes simultaneously.
★ Developed and deployed a Real-Time Traffic Light Detection model with 59% mAP for the autonomous driving shuttle.
★ Enhanced the speed of existing C++ based perception algorithms by 40X using CUDA C++.
Integrated a RADAR to Novus Aware device to add a Forward Collision Warning feature. Created a Qt-based GUI tool to visualize objects from Radar, Improved collision warning to be robust using Bayesian filters.
★ Improved driving data collection pipeline using AWS IoT to be 60% more cost-effective.
Constructed a real-time web streaming feature (WebRTC) for Novus Aware device to stream driving videos to clients.
★ Prototyped and devised 6 camera Surround View System (SVS) for commercial vehicles.
Team:- Mobile Robotics
★ Built a timed mission feature so that the robot could perform a specified task at a given time and integrated it with UI.
★ Improved robot localization and loop closure by fusing an IMU using EKF to existing RTAB-Map localization.
★ Created an AR Tag detection algorithm to enable fiducial marker-based navigation for mobile robots.
Designed various UI features for Robot fleet UI to enhance the user experience of clients using Ros2djs.
