DETECT→TRACK→ROAD PLANE→RANGE
LANE LOCK
Frame 0/— · lane lock — of pass
Embedded perception / Retrofit ADAS
RETROVISION
Modern road awareness for cars that predate it. Below is a real perception pass over 23.5 seconds of dashcam footage — lane fit, object tracks and ranges, all measured, then replayed frame for frame.
DETECT YOLO11-S / ByteTrack
LANE Bird's-eye fit
RANGE Ground-plane homography
Forward perception
I‑405 · 07/10/2024 · 1080P30
0.00 / 23.50
Road plane / metric
0 – 55 M
SCANNING
No objects in field of view.
No lead vehicle in the ego lane.
0.0 mph · headway —
Source clip: Rove R2-4K sample footage, seconds 6.0–29.5, desaturated for
legibility. Detection and lane fitting run offline in
tools/retrovision_track.py; the browser replays the result and
draws it. Ranges are monocular — recovered by projecting each object's
ground contact point through the road-plane homography — not LiDAR.
The premise
The car does not need a new dashboard. It needs better information.
RetroVision keeps the driver-facing layer intentionally quiet. It elevates only what changes a decision: lane position, the closest forward object, its class, and the level of caution required.
The visual language borrows from machine vision—not entertainment— so every mark has a measurable source behind it.
System architecture
Two chips. One clear contract.
Sensor work stays separate from display work, keeping perception responsive and the driver interface predictable.
Sense
Camera + LiDAR
Semantic objects from the camera. Absolute forward distance from the TFmini-S.
Fuse
ESP32-C3
Normalizes detections, range state, confidence, and the current planner decision.
{"class":2,"range":18.4}
Display
ESP32-S3 + LVGL
Renders the driving scene and turns the fused packet into glanceable warnings.
Say what it is.
Cars, pedestrians, traffic lights, and signs stay semantically distinct instead of becoming anonymous obstacles.
Show how close.
LiDAR gives the display a grounded forward distance even when camera confidence shifts frame to frame.
Surface the state.
Clear, caution, yield, and stop states remain visible—along with the evidence that produced them.