Astra Perception is a high-performance, distributed Advanced Driver Assistance System (ADAS) specifically engineered for the complexities of unstructured road environments. Built with a Cyber-Minimalist aesthetic, it combines real-time computer vision with intelligent threat analysis to provide a "Glass Cockpit" experience for any vehicle.
Unlike traditional ADAS that relies on perfect lane markings, Astra uses Multi-Scale Road Surface Modeling. It excels on roads with faded markings, potholes, and varied paving by identifying the physical drivable surface rather than just lines.
To solve the issue of detection flickering, the system features a Hallucination Memory. It maintains object persistence for up to 3 frames if the AI momentarily loses visibility, ensuring rock-solid and stable safety warnings.
The system utilizes a modular automotive architecture:
- Vision Engine (AI Lens): Manages high-frequency sensor data and heavy AI inference on the GPU.
- Control ECU (Cockpit): A standalone terminal that receives perception packets via low-latency UDP to manage vehicle logic and trajectory planning.
- Small-Object Detection: Powered by YOLOv8 Small for superior tracking of distant motorcycles, pedestrians, and obstacles.
- Dynamic Digital HUD: A 30 FPS persistent drivable area overlay featuring pulsing grid-pattern caching for zero CPU overhead.
- Surface Hazard Analysis: Adaptive Gaussian filtering to pinpoint potholes and speed bumps, cross-referenced with the drivable area mask.
- Trajectory Smoothing: Integrated Kalman Filters for precise distance and velocity prediction.
- Predictive Intent: Automatic detection of "Cut-In" maneuvers from the side before they enter the vehicle's direct path.
- Priority Resolution: A smart logic engine that isolates the single most critical threat to prevent alert fatigue.
- Virtual Actuators: Real-time PID controllers for simulated steering and acceleration/braking response.
- Tactical Radar: A 2D spatial visualization dashboard featuring motion trails for all tracked objects.
- Inference: Ultralytics YOLOv8 (GPU / CUDA / FP16)
- Tracking: Kalman-Filtered Centroid Tracking with Persistence Memory
- Vision Logic: OpenCV (LAB Color Space, L1 Manhattan Distance Optimization)
- Interface: PyQt6 with Cyber-Minimalist Glass-Panel CSS
- Concurrency: Asynchronous Multi-threaded Producer-Consumer Pipeline
- Python 3.12 or higher
- NVIDIA GPU with CUDA 12 support (Automatic high-performance CPU fallback)
git clone https://github.com/Sherin-SEF-AI/Astra-Perception.git
cd Astra-Perception
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txtThe suite provides three distinct operational modes accessible via the application menu:
- Vision Engine: The primary perception and AI analysis module.
- Control ECU: The standalone radar and cockpit telemetry dashboard.
- Integrated Stack: Simultaneous launch of the complete perception and control system.
Sherin Joseph Roy Professional AI Engineer specializing in Autonomous Systems and Computer Vision.
Distributed under the MIT License. See LICENSE for more information.