🚗 AutoVision: Semantic Segmentation & Object Detection for Autonomous Vehicles

AutoVision is a deep learning-based computer vision project designed to enhance perception in autonomous vehicles. It covers semantic segmentation and object detection using both camera images and LiDAR-based data.

🔍 Project Highlights

• Semantic Segmentation

  • Dataset: Lyft Udacity Semantic Segmentation Dataset
  • Models: U-Net, Attention U-Net, Inception U-Net, Residual U-Net
  • Best Performance: 97.2% accuracy with Inception U-Net
  • Evaluation Metrics: IoU, Dice Coefficient

• Object Detection

  • Datasets: LiDAR-based dataset from Kaggle (3D Object Detection for Autonomous Vehicles)
  • Models: YOLOv3 (custom Keras implementation), YOLOv8
  • Focus: Accurate 3D object localization using LiDAR and camera data

⚙️ Tech Stack

• Python, TensorFlow, Keras, OpenCV
• U-Net and YOLO architectures
• LiDAR and RGB image data
• Jupyter Notebooks for training and visualization

📊 Results

• Achieved high segmentation accuracy (IoU > 0.95) using advanced U-Net variants
• Detected vehicles and obstacles with high precision using YOLO models

🚀 Future Improvements

• Integrate panoptic segmentation
• Fuse LiDAR and RGB data for joint learning
• Deploy lightweight versions for edge devices

🧠 Author

Vishal – AI/ML & Computer Vision Enthusiast Refer in my kaggle profile- https://www.kaggle.com/code/helloworld349759/autovision