NXP Car Controller end-to-end pipeline using imitation learning

Developed a Deep-learning based framework for visual navigation using Imitation Learning in an urban environment laid with obstacles such as pedestrians, barricades and other cars. Deployed End-to-End CNN model with regularization technique with ROS2 in a simulated environment.

For Object Detection

• Significantly improved accuracy from a very small dataset of 1000 images for 5 categories, by creating custom dataset with image augmentation and data scraping; And smart pre-training of YOLOv4 Tiny. Creating NXP car control using convolutional neural networks

Data

• Data Files: https://drive.google.com/drive/folders/1TLdkgaS9dncPFjXhM6tqCrHvLh1JJjY7?usp=sharing

Training Procedure

• First run the environment simulator
• Then run the data generation file to capture data from the simulator
• Control the simulator with the joystick to create data

python3   train_data_gen.py

• After enough data is generated we can move on to train our model using:

python3   train.py

Deployment

• After trainning the learned controller can be deployed by using

python3   controller.py