yzfzzz/depth-detect

What it solves

This project implements an intelligent rear-view mirror system for bicycles. It addresses the need for real-time hazard detection by combining object detection with depth estimation to identify "fast-approaching" vehicles. It specifically solves the challenge of running these computationally expensive models on resource-constrained embedded platforms like NVIDIA Jetson.

How it works

The system fuses YOLO for object detection and various depth estimation models (such as Depth-Anything, Lite-Mono, or YOLO-Depth). It determines if a target is dangerous by monitoring if the depth is decreasing while the bounding box height is increasing over time.

To optimize performance, it employs three scheduling modes:

• Serial: Detection and depth estimation run sequentially.
• Overlap: Both models run asynchronously using CUDA streams to maximize GPU utilization.
• Stagger: Models are scheduled at different intervals (e.g., detection every frame, depth every 3rd frame), reusing previous results to increase overall FPS.

It uses BYTETracker for multi-object tracking and reports alerts via JSON over TCP. The entire pipeline is accelerated using TensorRT and CUDA.

Who it’s for

• Developers building smart vehicle safety systems or ADAS (Advanced Driver Assistance Systems).
• Engineers working with embedded AI on NVIDIA Jetson (TX2, Nano) or x86 Linux platforms.
• Researchers interested in real-time fusion of 2D detection and monocular/binocular depth estimation.

Highlights

• Flexible Scheduling: Supports sync, async overlap, and staggered inference to balance latency and throughput.
• Embedded Optimized: Fully compatible with Jetson TX2/Nano and x86, with TensorRT FP16 acceleration.
• Danger Detection: Implements a custom "approach" logic based on depth change rates and bounding box growth.
• Robust IO: Features an asynchronous disk-writing buffer to prevent IO bottlenecks from slowing down the inference pipeline.