Launcher

Type to filter results. Use arrow keys to navigate, Enter to select.

Preferences

You've reached the end of the results

Select Enter
Navigate Arrow down Arrow up
Close Esc

Obstacle Avoidance Car with Monocular Depth Estimation

A graduation project autonomous car that steers using real-time monocular depth estimation running on a Raspberry Pi.

Project Overview

My bachelor’s graduation project — titled “Deep Learning and Its Application” — an autonomous four-wheel-drive car that determines its driving direction purely from a single camera feed, using the MiDaS monocular depth estimation model to infer scene depth without stereo cameras or LiDAR.

Car Source Code | MiDaS C++ Inference Example | Thesis (Turkish) | Demo Video

Technical Details

  • The MiDaS v2.1-small depth estimation model — run through TensorFlow Lite and OpenCV, entirely in C++ — processes the live camera feed to produce a depth map.
  • A simple detection algorithm extracts a layer of that depth estimation into a binary image, then decides whether to steer left or right by analyzing the distribution of pixels across it.
  • The model and control algorithms run on a Raspberry Pi 4 as the onboard single-board computer, with an Arduino handling low-level hardware control (motors and steering).
  • Combines Python (for AI model inference) and C++ (for hardware control and the final MiDaS deployment) in a single pipeline.

Results

Benchmarked on a Raspberry Pi 4 Model B (4 GB RAM, 64-bit Ubuntu 22.04, no overclocking): roughly 2.3 FPS using all cores, and 0.91 FPS single-threaded.

MiDaS depth map output next to the source camera frame, running live on the Raspberry Pi A second MiDaS depth map example, showing the depth output and camera overlap windows on the Raspberry Pi desktop

Outcome

The project was well received by my thesis instructor as a demonstration of practical, real-time AI deployment on constrained embedded hardware.

References