Autonomous Driving V2X Preprocessing Pipeline

Master's Thesis Research

This repository contains the data engineering pipeline required to convert the Waymo Open Motion Dataset into multi-dimensional, deep-learning-ready tensors optimized for Controllable Diffusion Models (MotionDiffuser).

Features

1. Raw Ingestion: Deserializes Google Protocol Buffers (.tfrecord).
2. Ego-Centric Normalization: Translates global GPS coordinate systems to an Ego-centric reference frame (0,0), applying rotational matrices to align agent trajectory vectors.
3. Tensor Engineering: Pads variable traffic scenarios into uniform [64, 91, 6] tensors, integrating binary Valid Masks for Transformer attention layers.

Execution

To generate the final .npy tensors:

cd src/
python3 build_tensors.py