Self-Driving Vehicular Project

Team: Aaron Cruz [UG], Arya Shetty [UG], Brandon Cheng [UG], Tommy Chu [UG], Erik Nießen [HS], Siddarth Malhotra [HS]

Advisors: Ivan Seskar and Jennifer Shane

---

Project Description & Goals:
Build and train miniature autonomous cars to drive in a miniature city.
RASCAL (Robotic Autonomous Scale Car for Adaptive Learning):

• Using the car sensors, offload image and control data onto a cloud server.
• This server will train a neural network for our vehicle to use in order to drive autonomously given camera image data.
  

Technologies: ROS (Robot Operating System), Pytorch
​https://gitlab.orbit-lab.org/self-driving-car-2023/

Documentation: ​https://docs.google.com/document/d/1KWIyvGMnB2nboosv068qp-X2BbNm9y4tUnWLzKbz6bI/edit#heading=h.qhh6ahjpnjdi

---

Week 1: 5/28 - 5/30

​Week 1 Slides

Progress:

• Got familiarized with past summer's work:​GitLab, RASCAL setup, ​Poster
• Debugged issue with RASCAL's pure pursuit

---

Week 2: 6/3 - 6/6

​Week 2 Slides

Progress:

• Setup X11 forwarding for GUI applications through SSH
• Visual odometry using Realsense Camera and ​rtabmap
• Streamlined data pipeline that processes bag data (car camera + control data) into .mp4 video
• Detected ARUCO markers from a given image using Python & OpenCV libraries
• Setup Intersection server (node with GPU)
• Developed PyTorch MNIST model
• Trained "yellow thing" neural network
• Line up perspective drawing with camera to determine FOV

---

Week 3: 6/10 - 6/13

​Week 3 Slides

Progress:

• Created web display assassin to eliminate web server when closing ROS
• Tested "yellow thing" model, great results
• SSHFS setup
• Calibrated Realsense camera
• Created "snap picture" button on web display for convenience
• Developed python script to detect ARUCO marker and estimate camera position
• Tested point cloud mapping with rtabmap
• Attempt sensor fusion with encoder odometry and visual odometry
• Data augmentation to artificially generate new camera perspectives from existing images

---

Week 4: 6/17 - 6/20

​Week 4 Slides

Progress:

• Refined aruco marker detection for more accurate car pose estimation
• Trained model with video instead of images
• Improved data pipeline from car sensors to server
• Refined data augmentation to simulate new camera perspectives
• Added more data visualization (Replayer) to display steering curve, path, and images to web server

---

Week 5: 6/24 - 6/27

​Week 5 Slides

Progress:

• Aruco Marker Detection now updates car position within XY plane. Finished self-calibration system
• Addressed normalization and cropping problems
• Introduced Grad-CAM heat map
• Resolved Python version mismatch issue
• Visualized training data bias through histogram
• Smoothed data to reduce inconsistency in training data
• Simulation Camera - skews closest image to simulate new view
• Added web display improvements - search commands, controller keybinds

---

Week 6: 7/1 - 7/3

​Week 6 Slides

Progress:

• Curvature interpolation and calibration
• Trained model using double orange barriers
• Training batches are less 0 biased
• Data augmentation blur artifacts during NN normalization - solid fill
• Simulated driving with ML model and image skewing
• Improvements to Aruco detection: code refactoring, functionality for multiple Arucos

---

Week 7: 7/8 - 7/11

​Week 7 Slides

Progress:

• Rascal Simulator on server
• Updated CAD models for rascal
• YOLO (You Only Look Once) Object detection and distance
• City training data using a constrained path
• Automated testing and evaluation for models
• Aruco Detection Application: integrated into city, testing with pursuit loop

---

Week 8: 7/15 - 7/18

​Week 8 Slides

Progress:

• Spline processing & model, calculates future trajectory based on best fit spline
• Setting up Virtual Machine & documentation
• YOLO - Cone & barrier detection
• Hyperparameter optimization for model
• Automation, testing, and adjustments for aruco detection

---

Week 9: 7/22 - 7/25

​Week 9 Slides

Progress:

• Data loader improvements: separate replaying and data loader, delete data, load multiple sessions at once
• More city training data - lane correction, crash avoidance, orange blocks to limit path
• Efficient automation for city training
• YOLO: traffic light detection, obtaining speed vs accuracy
• Continuing hyperparameter optimization

---

Week 10: 7/29 - 8/1

​Week 10 Slides

Progress:

• Quality of life changes for data collection: camera stream refresh, bag size output
• New hardware for new car: 3D printed parts, fisheye camera, wireless controller
• Fisheye camera: Calibration and undistorting images
• YOLO: custom dataset training, red light finder
• Hyperparameter optimization improvements
• Training city data with the new fisheye camera!

---

Final Week: 8/5 - 8/7

​Open House Presentation

​Final Poster

---