Coursework

As part of my Autonomous Vehicle and Machine Learning course, I worked with the MX Car Kit, equipped with stereo cameras, Lidar, and a NVIDIA Jetson Nano, to deploy real-time perception and control models. The development environment included ROS2 Humble, Docker, and Foxglove for visualization and debugging.

Midterm Project: Traffic Sign Detection and Rule-Based Control on MX Car Platform

For the midterm, I developed a traffic-aware control system using a YOLO-based object detection model trained on a custom dataset with the following object classes:

• Lane lines

• Vehicles

• Stop sign

• Traffic lights (red, yellow, green)

• Speed signs (2 m/s and 3 m/s)

Key features of the project:

• Built a ROS2 workspace to process stereo camera input and generate real-time object classifications and distance estimates.

• Developed custom ROS2 nodes to control vehicle behavior in response to detected objects and traffic conditions.

• Implemented rule-based decision logic to follow lane lines, stop at stop signs and red lights, and adjust speed based on visible signage.

Final Project: Multi-Output PilotNet for Adaptive Cruise Control using Behavior Cloning

Developed a multi-output deep learning model based on PilotNet architecture to enable adaptive cruise control on an autonomous vehicle using behavior cloning. The system was trained on synchronized sensor data collected from ROS bagfiles, including ground truth steering angles and a proxy for following speed.

• Due to inaccuracies in the recorded following speed data, I implemented a custom data pipeline:

• Applied a YOLO object detection model to identify leading vehicles in camera frames.

• Extracted corresponding Lidar-based distance measurements to those vehicles.

• Filtered and interpolated the Lidar data to generate a clean distance signal.

• Derived following speed using a kinematic model based on the change in distance over time.

Model training was conducted using Google Colab with GPU acceleration. Training progress and performance metrics (e.g., loss curves, accuracy) were tracked using Weights & Biases (W&B) for experiment management and visualization.