Self-Driving-Cars - Object-Tracking

Please check Slides and Project Report for more details and figures

Sample Visualiztion

Point Cloud Sequence with 3D BBox tracked and its Corresponding Video Sequence

This codebase is built on top of https://github.com/Ghostish/Open3DSOT, with references from https://github.com/HaozheQi/P2B and https://github.com/haooozi/OSP2B. Our Codebase was setup on GCP with 100gb storage for this project and NVIDIA L4 GPU(24GB RAM) and g2-standard-8 8vcpus 32GB RAM

Installation

environment setup

cd 3dsot conda create -n myenv python=3.8 conda activate myenv

install pytorch with cuda support

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117

install requirements

pip install -r requirement.txt

Download KITTI Velodyne dataset, calib and label_02 files and unzip them and place them as follows

--> [calib]
    --> {0000-0020}.txt
--> [label_02]
    --> {0000-0020}.txt
--> [velodyne]
    --> [0000-0020] folders with velodynes .bin files

Training the model on Car category

CUDA_VISIBLE_DEVICES=0 python main.py --cfg cfgs/P2B_Car.yaml --batch_size 64 --epoch 60 --category_name Car

Testing model on Car category

python main.py --cfg cfgs/P2B_Car.yaml --checkpoint ./ckpts/car-31.ckpt --test --category_name Car

To run tensorboard after training

pip install tensorboard
tensorboard --logdir=./ --port=6006

trained model checkpoints present in 3dsot/ckpts/