HazDepth:

Real-Time Hazard Detection with Depth Estimation

A monocular computer vision pipeline that tells users not just what is nearby, but how far away it is. No depth sensor required at inference. Trained specially for hazardous objects.

The Problem HazDepth Solves

Cameras see objects. Not distances.

Smart glasses for the visually impaired can detect a kitchen knife — but without knowing it is 40 cm away, that information is functionally useless. Existing assistive vision systems lack the spatial awareness needed for real-world safety-critical navigation.

Why existing solutions fall short

Stereo & LiDAR cameras

Expensive, bulky - not viable for wearable assistive tech

Monocular depth alone

No semantic hazard classification; cannot name the object or its class

Object detection alone

Bounding boxes with no distance, insufficient for navigation safety

How HazDepth Works

A tightly integrated two-stage pipeline combining state-of-the-art monocular depth estimation with robust object detection - calibrated against ground-truth depth hardware, deployed on a single RGB camera.

At inference, only one RGB camera is required. The system was calibrated against an Intel RealSense D435i to ground depth estimates in real-world metric units — producing outputs like "Knife — 40 cm ahead" at 15.5 FPS on a single GPU.

The Dataset: 2,000 Hand-Verified RGB-D Pairs

Why fully manual annotation?

Automated labelling tools — including YOLO and large language models — failed on 88–90% of safety-critical objects in preliminary testing. This failure rate was itself an original finding, confirming that hazard-class objects require human verification for reliable ground truth.

2,000

RGB-D image pairs, manually verified

7

Hazard object classes annotated

3

Distinct indoor environments

100%

Hand-annotated by researchers

Performance Against Competing Methods

HazDepth outperforms all comparable monocular depth baselines on safety-critical indoor scenes. The fusion architecture closes the gap between depth estimation quality and actionable hazard awareness.

MonoDepth2

No object-level semantic output; depth accuracy insufficient for close-range hazard detection

BTS & AdaBins

Strong outdoor depth estimation but no hazard class fusion; not designed for wearable assistive contexts

MiDaS v3.1

Relative depth only — cannot produce metric distances without separate calibration and object detection

ZoeDepth

Metric depth capable but operates independently of object detection; no spatial hazard output

Made with