Autonomous Vehicle Perception System
Multi-camera ADAS platform for autonomous vehicle development
- Status
- Research
- Version
- No tagged release
- Platform
- Desktop app
- License
- Not specified
Provides a single desktop platform for autonomous vehicle development that brings multi-camera capture, perception, safety monitoring, planning and testing together in one modular application.
How it is put together.
Autonomous Vehicle Perception System is a PyQt6 desktop application that captures from up to four USB cameras in parallel threads and runs a perception pipeline on them: classical lane detection, YOLOv8 object detection, centroid tracking, multi-camera fusion and a bird's-eye-view display. Later versions add safety warnings (forward collision, lane departure, blind spot), driver monitoring, recording and playback, calibration, path and motion planning, vehicle control models, a V2X module and a scenario testing framework. It was built in phases, each documented in the README and CHANGELOG.
- 1
Camera subsystem
Capture threads, multi-camera manager, frame synchronizer, calibration
camera/ - 2
Perception
Lane, object, tracking, fusion, segmentation, depth, signs and lights, visual odometry
perception/ - 3
Safety systems
Collision, lane departure, blind spot, driver monitoring, safety scorer
safety/ - 4
Planning
Path, route, motion and behavior planners; parking assist
planning/ - 5
Recording and playback
Recording manager, video and metadata writers, playback
recording/ - 6
User interface
Main window, camera widgets, telemetry dashboard, calibration wizard
ui/ - 7
Visualization
BEV generator, minimap, overlay renderer
visualization/
Specification.
- Cameras
- Up to 4 UVC USB cameras, one QThread each, with frame synchronization and auto-reconnect
- Perception
- Canny/Hough lane detection with polynomial fit; YOLOv8n detection; centroid tracking; multi-camera fusion; BEV view
- Safety
- Forward collision (TTC), lane departure (TTLC), blind spot warning, driver monitoring, safety score 0-100
- Planning and control
- A*, RRT and Frenet planning; kinematic/dynamic bicycle models; PID, Stanley and Pure Pursuit controllers
- Recording
- Synchronized multi-camera MP4 recording with JSONL metadata; playback; CSV/JSON/video export
- Calibration
- Chessboard calibration wizard for intrinsics and distortion, saved as JSON
- Benchmark thresholds
- pytest-benchmark tests assert lane detection < 50 ms at 640x480, tracking < 10 ms for 10 objects, pipeline > 15 FPS
- Runs on
- PyQt6 desktop application launched with python3 main.py and fed by local USB cameras
More systems.
Real-time, on-device driver assistance for any Android phone.
Perception and ADAS · LiveDashLabTrain object detectors and visual models on dashcam footage without labelling any of it
Next systemAutoSim ProAutonomous vehicle simulation platform with vehicle physics, AI behaviours and analytics