AI-Powered Object Detection and Risk Assessment System

Project Overview

ORAMA is an AI-based system designed to enhance vehicle safety in adverse weather conditions, specifically foggy and rainy scenarios. It improves object detection and provides real-time collision risk assessment. Developed as a Minimum Viable Product (MVP) for a hackathon, ORAMA leverages:

• YOLOv8 for object detection.
• Weather-adaptive preprocessing for visibility enhancement.
• A heuristic-based risk assessment algorithm to identify potential collisions.

The system processes video input, detects objects, and highlights collision risks with visual alerts, achieving real-time performance at ~29 FPS.

Key Features

• Weather-Adaptive Preprocessing: Dynamically adjusts for fog and night conditions using dehazing and brightness/contrast enhancement.
• Object Detection: Utilizes YOLOv8 to detect collision-relevant objects (cars, trucks, pedestrians) in low-visibility scenarios.
• Risk Assessment: Estimates object distance and direction, calculating a risk percentage to identify potential collisions.
• Visual Alerts: Highlights risky objects with red bounding boxes and displays the risk percentage (e.g., Risk: 45.3%).
• Real-Time Performance: Processes frames at ~34.5ms per frame (~29 FPS), suitable for edge deployment.

Vision

ORAMA aims to become a Software-as-a-Service (SaaS) platform for autonomous vehicle manufacturers, providing real-time object detection and risk assessment as a standardized solution. Future plans include:

• Ethical data collection from vehicle cameras (with user consent).
• Integration with additional datasets for diverse driving conditions.
• Deployment on high-performance edge devices for real-world validation.

Problem Statement

Visibility is significantly reduced in foggy and rainy conditions, making traditional object detection methods unreliable. ORAMA addresses this by:

• Selecting and processing relevant datasets.
• Developing deep learning models for detection.
• Implementing adaptive preprocessing techniques.
• Evaluating performance and risk assessment models.
• Designing real-time alert mechanisms for drivers.

Repository Structure

ORAMA/
├── orama/
│   ├── inference.py              # Main script for inference and risk assessment
│   ├── output/
│   │   ├── inference_log_<timestamp>.log  # Log files
│   │   └── output_inference_<timestamp>.mp4  # Output videos
│   ├── runs/
│   │   └── train3/
│   │       └── weights/
│   │           └── best.pt      # Trained YOLOv8 model weights
│   └── (evaluation outputs: confusion_matrix_normal.png, F1_curve.png, etc.)
├── yolo_train/                   # Directory for training-related files
├── convert_boco_to_yolo.py       # Script to convert BDD100K dataset to YOLO format
├── evaluate.py                   # Script to evaluate model performance
├── filter_bdd100k.py             # Script to filter BDD100K dataset
├── train_yolo.py                 # Script to train YOLOv8 model
├── video.mp4                     # Input video for inference
└── README.md                     # This file

Setup Instructions

Prerequisites

• Python 3.8+
• Operating System: Tested on Ubuntu 20.04; should work on Windows and macOS with minor adjustments.
• Hardware: GPU recommended for faster inference (e.g., NVIDIA GPU with CUDA support).

Installation

Clone the Repository

git clone https://github.com/Aaris03Khan/orama.git
cd orama

Set Up a Virtual Environment (Recommended)

python3 -m venv orama.env
source orama.env/bin/activate  # On Windows: orama.env\Scripts\activate

Install Dependencies

pip install ultralytics opencv-python numpy

Download the Model Weights

• Place the trained YOLOv8 model weights (best.pt) in orama/runs/train3/weights/.
• If not included, train the model using train_yolo.py (see Training section).

Prepare the Input Video

• Place your input video as video.mp4 in the ORAMA/ directory.
• Alternatively, modify the video path in inference.py.

Usage Instructions

Running the Inference Script

Navigate to the ORAMA directory:

cd orama

Run the script:

python3 inference.py

Viewing the Output

The output video will be saved as:

output/output_inference_<timestamp>.mp4

Logs are saved in output/inference_log_<timestamp>.log.

Results and Performance

Output Video

• Left Side: Raw video with frame counter.
• Right Side: Processed video with detected objects (green bounding boxes) and collision risks (red bounding boxes with risk percentage).

Performance Metrics

• Detection Accuracy:
  • mAP: 0.472 (normal conditions), 0.274 (foggy conditions)
  • IoU: 0.65 (normal), 0.55 (foggy)
• Real-Time Performance: 34.5ms per frame (29 FPS), suitable for edge deployment.

Future Work

• Dataset Expansion: Add Foggy Cityscapes, KITTI, and synthetic data.
• Preprocessing: Integrate deep learning-based dehazing and rain removal (e.g., U-Net).
• Risk Assessment: Implement velocity and trajectory analysis using object tracking (DeepSORT).
• Alerts: Introduce audio alerts and dashboard notifications.
• Vehicle Assistance: Simulate automatic braking and lane departure warnings.
• Edge Deployment: Optimize inference on NVIDIA Jetson Nano.

---

ORAMA is a scalable AI-powered system that enhances road safety in challenging weather conditions. Future developments will focus on improved accuracy, ethical data collection, and seamless integration with autonomous vehicle platforms.