VisionDriveX — Explainable Autonomous Driving Perception System 🚗🧠

VisionDriveX is a multi-task autonomous driving perception system that performs traffic-sign classification, stop-sign detection, and lane segmentation, with explainable AI (Grad-CAM) to visualize model attention. Designed for research, demo presentations, and real-time AV perception prototypes.

🔥Features

Traffic Sign Classification (GTSRB — 43 classes) Stop-Sign Detection using Faster R-CNN Lane Segmentation (binary lane mask) Real-Time Pipeline (webcam input, ESC to exit) Explainability using Grad-CAM overlays Confusion Matrix + Evaluation Scripts Modular folder structure for training, inference, and deployment

⚙️ Installation 1️⃣ Create Virtual Environment py -m venv .venv .venv\Scripts\Activate.ps1

2️⃣ Install Packages py -m pip install --upgrade pip py -m pip install -r requirements.txt

🧠 GTSRB Human Labels File

Generate class-name mapping: py scripts\make_labels.py

Creates: data/gtsrb/labels.json

🎥 Real-Time Demo (Webcam)

Run the fast perception pipeline:

py -m src.realtime.realtime_fast --device cpu --cam 0 ^ --cls_weights checkpoints/classifier_resnet18_epoch3.pth ^ --det_weights checkpoints/det/fasterrcnn_epoch3.pth ^ --seg_weights checkpoints/seg/seg_epoch3.pth ^ --img_size 128 --num_classes 43

Real-time outputs include: CLS: 25 (0.92) → class index + confidence Green bounding boxes → Stop-sign detection Lane mask overlay FPS counter

Press ESC to exit.

🧪 Model Training 1️⃣ Train Traffic-Sign Classifier py -m src.train.train_single_task --task classification ^ --data data/gtsrb/train --num_classes 43 --epochs 30 ^ --batch_size 16 --img_size 224 --backbone resnet18 ^ --checkpoint_dir checkpoints --device cpu

2️⃣ Train Lane Segmentation py -m src.train.train_segmentation --images data/tusimple/images ^ --masks data/tusimple/masks --epochs 20 --batch_size 4 ^ --checkpoint_dir checkpoints/seg --device cpu

3️⃣ Train Stop-Sign Detector (COCO format) py -m src.train.train_detection_coco --images data/custom_stop/images ^ --ann data/custom_stop/annotations.json --num_classes 2 ^ --epochs 10 --checkpoint_dir checkpoints/det --device cpu

📊 Evaluation Confusion Matrix + Classification Report py scripts\eval_classifier_cm.py ^ --checkpoint checkpoints/classifier_resnet18_epoch3.pth ^ --data_root data/gtsrb/val --img_size 224 --device cpu

Outputs saved to:

outputs/confusion_matrix.png outputs/per_class_accuracy.csv

🔍 Explainability (Grad-CAM)

Generate Grad-CAM for a single input image:

py src/explainability/gradcam_demo.py ^ --image data/gtsrb/val/14/00001.png ^ --weights checkpoints/classifier_resnet18_epoch3.pth ^ --out outputs/gradcam.png ^ --img_size 224 --device cpu

🌐 Optional Streamlit Demo

Run a simple web app: streamlit run streamlit_app.py

🛣️ Demonstration Ideas for Presentations

Show webcam feed with: Phone screen showing traffic-sign PNGs STOP-sign printed image Drawn white lines for lane segmentation Side-by-side Grad-CAM overlay (what the model looks at)

Explain: Multi-task perception Real-time inference Explainability and safety relevance

📄 License MIT License © 2025 Mouna C

👤 Author VisionDriveX — Explainable Autonomous Driving Perception System Created by: Mouna C GitHub: https://github.com/labcodesss