- Cars are equipped with cameras.
- We want to build ADAS.
- Use cameras on board to assist the human driver or to self-drive
- As an ADS, it needs to plan ahead so that it can move safely in RMC.
- As a planning engineer, I want to know where the free space is on the road.
- As an ADS, it needs to keep lane so that it can drive safely.
- As a planning engineer, I want to know where the lane lines are.
- As an ADS, it needs to change lane so that it can go to destination.
- As a planning engineer, I want to know where the lane lines are.
- ADS: Automated Driving System
- RMC: Minimum Risk Condition
- (Main) Detect where the free space is on the road As Good As Tesla
- Color the free space of the road on the image by "image segmentation techinque"
- (Optional) Detect lane lines As Good As Tesla
- Draw continous lines/curves which indicates lane lines
- Research state-of-the-art method to detect the free space on the road
- Decide the proper image segmentation techinque for the project among many
- Use dataset obtained from real car
- Build ML model and train/eval/test
- Implement Custom Pytorch Dataset
- Implement Custom Pytorch Dataloader
- Implement Custom Pytorch module
- Get images from dataset
- Annotate date
- Augment dataset
- Implement unit tests
- Implement MLOps pipeline
- Carry out unit tests
- Turn Pytorch model to ONNX model
- Turn ONNX model to TensorRT model
- Implement integration test
- Carry out integration test
- Deploy to Jetson TX2
-
W-net (Fully Unsupervised Image Segmentation)
-
FCN-8 (Fully Convolutional Networks)
- Evan Shelhamer, et al., Fully Convolutional Networks for Semantic Segmentation (2016)
- Jonathan Long, et al., Fully Convolutional Networks for Semantic Segmentation (2015)
- Luis C. García-Peraza-Herrera, Real-Time Segmentation of Non-Rigid Surgical Tools based on Deep Learning and Tracking (2020)
- Karen Simonyan and Andrew Zisserman, Very Deep Convolutional Networks for Large-Scale Image Recognition (2015)
-
SCNN (Spacial CNN)