drive-guardian

Drive Guardian

An open-source Advanced Driver Assistance System (ADAS) project, re-architectured for high scalability and adaptability from powerful backend servers to resource-constrained edge devices (like the Raspberry Pi).

Architecture

• Core: Hardware abstraction layer & profile-based configuration loading.
• Inference Engine: Executes deep-learning logic (Object detection, segmentation) independent of the UI. Supported through YOLO / ONNX runtime.
• Desktop App: Native Python UI built with CustomTkinter.
• Edge Communications: IoT integrations for MQTT or CAN-bus reading.
• ML Pipeline: Experimental notebooks and scripts for model training and exporting.

YOLOv8 Throughput Optimization

The PyTorch YOLO runner now supports two YOLOv8-focused performance toggles through profile YAML files:

• YOLO_ENABLE_FUSE: Fuses convolution + batch norm layers to reduce inference overhead.
• YOLO_ENABLE_HALF: Uses FP16 on CUDA for higher throughput.
• YOLO_ENABLE_DCP_DEHAZE: Applies Dark Channel Prior (DCP) dehazing before YOLO inference for fog/haze robustness.

Example profile values:

YOLO_ENABLE_FUSE: true
YOLO_ENABLE_HALF: true
YOLO_ENABLE_DCP_DEHAZE: true

Notes:

• These options are applied when using a PyTorch .pt model via YoloRunner.
• ONNX runtime paths are not affected by these toggles.
• The desktop app also exposes a live YOLO DCP Dehaze switch so you can turn the preprocessing on or off without restarting.

Dark Channel Prior (DCP) Dehazing

DCP is an image restoration method that estimates haze thickness from low-intensity local patches (the dark channel), estimates atmospheric light, and reconstructs a clearer scene before detection. In Drive Guardian this is implemented as a lightweight pre-processing function to improve object visibility in hazy conditions.