SFU MSE capstone project Jan 2022-Aug 2022
Supervisor: Dr. Ahmad Rad
Xilun Zhang
Artur Shadnik
Delin Ma
Zhen Li
Jaehong Kim
-
CARLA Simulation with built-in Stereo Camera
-
CARLA Simulation with HIL LiDAR
-
OpenCV Lane Navigation (Urban road environment)
-
PilotNet Lane Navigation (Urban road environment)
-
PilotNet Lane Navigation with transfer learning (SFU Surrey Galleria 4)
-
Object Classification using ZED Camera (Yolo-v4-tiny)
This project aims to improve the functionalities of a level-2 automated vehicle prototype and algorithms. The expected outcome is level-4 automation in ODD. The expected system architecture is shown in the figure below. The main design criteria of this project are environment perception, high-level control (motion planning) and low-level vehicle control (lateral, longitudinal and speed).
For more details, please refer to our final report or leave an email at xza213@sfu.ca
- Sensors
• The ZED Camera has bad hardware connection, which might lead to danger actions due to lose of camera data during driving.
• The RPLiDAR can only generate 2D map. To have a better understanding of the environment, a LiDAR that maps 3D environment is required to detect objects lower than the ego vehicle.
• One extra camera should be added at the back of our ego vehicle to detect the states of approaching vehicle such as velocity and distance.
• Sensor Fusion algorithm should be added to combine stereo camera and Lidar to reduce variance and noise.
- Perception
• Broaden the YOLO dataset with images of all classes of objects.
• Broaden dataset to cater different driving environment such as urban road and indoor environment.
• Implemented Transfer Learning to our own network. Combine pre-trained network such as ImageNet and MobileNet with self-collected dataset.
• Ensure the perception system can detect incomplete objects for each class including two overlapping objects such as pedestrian holding a stop sign.
- Waypoints Tracking
• Behavior planners should take corresponding actions upon detecting traffic lights and speed signs.
• The system should provide better performance on object motion prediction. The current model we are using is constant velocity prediction model which is not realistic. A more comprehensive prediction model should be implemented such as using DNN and least square regression.
• Auto-generate waypoints using lane navigation model, SLAM or A*/RRT Search.
- Lane Tracking
• Filters need to be added to OpenCV to detect lanes under various weather conditions.
• For the deep learning approach, more dataset was required to achieve a better performance.
• Optimize the transfer learning approach or PilotNet structure to cater specific testing environment.
- Vehicle Controllers
• Output signals from the low-level controllers must be adequately mapped for the microcontroller.