README.md

Dynamic Uncertainty-Aware BEV Occupancy Grid

MAHE Hackathon 2026 - AI in Mobility - Problem Statement 3

Project Overview

A two-stage pipeline that converts front camera images into metrically accurate Bird's Eye View occupancy grids for Level 4 autonomous vehicles.

Results

Method	IoU Score
Baseline	0.0013
Traditional IPM	0.2747
Neural Network	0.4844
Best Sample	0.6008
Improvement	210x better

Model Architecture

• Stage 1: Real Inverse Perspective Mapping (IPM)
• Stage 2: Custom BEV Refinement CNN
• Encoder: Conv2D blocks (3->32->64->128)
• Decoder: ConvTranspose2D blocks (128->16->1)
• Parameters: 484,000+
• Loss: Binary Cross Entropy
• Optimizer: Adam

Dataset

nuScenes Mini Dataset

• 15 scenes used
• 404 total samples
• 28,444 LiDAR points per frame
• Resolution: 1600x900

Setup and Installation

pip install nuscenes-devkit torch opencv-python matplotlib scipy pyquaternion

How to Run

1. Open BEV_Hackathon.ipynb in Google Colab
2. Mount Google Drive
3. Run all cells sequentially
4. View results automatically

Live Demo

Click here for Live Streamlit Demo

Results Images

Tech Stack

• Python 3.12
• PyTorch 2.x
• OpenCV
• nuScenes devkit
• Google Colab T4 GPU
• Streamlit (deployment)