Radar Vision System

2022-ECE, 22-PW-2nd, 22-V-HM

The objective of this project is to design a system that detects the speed of vehicles and logs their license plate numbers if they exceed a speed threshold

 

Team Members

Steven Horvath | Kaveesha Siribaddana | Yelizaveta Gadirova | Alexandros Sfikas |

Project Poster

Click on any image to enlarge.

 

Watch the Project Video

video player icon

Download the Project Summary

video player icon

Project Summary

 

Overview

There are various radar/video systems currently on the market that can detect speed and license plates. However, current radar/video systems lack modularity, are expensive, and are exclusive to the companies that use them. Our sponsor, Northrop Grumman, has tasked us to build a similar system that is affordable and uses off-the-shelf components. There is a need for a proof-of-concept system that is accurate to the company’s specifications.

Objectives

Our objectives were to create a modular speed trap, using the OPS242 radar, off-the-shelf parts and open-sourced code. Our machine learning model had to identify at least 50% of all license plates. Our system needed to save the license plate number and speed in a log file.

Approach

  • The design flow of the project was simple: Receive radar and video information, process them, and provide an appropriate output and log
  • Radar
    • OPS242 Doppler Radar
    • Receive serial data via USB
    • Python program interprets the data and applies a velocity threshold
  • Video
    • IMX477 Sensor, 1080p @60 FPS
    • Video data is processed into machine learning pipeline
  • Machine Learning
    • NVIDIA Jetson Nano as edge-computing device
    • Two different approaches: running TensorFlow model and running models with NVIDIA DeepStream SDK
      • After testing, we found that DeepStream was successful
    • Video from camera is processed through three cascading models; first, it detects a car in the image, then detect the license plate on that car, and finally detect the numbers on that plate
  • Speed and license Information is held in separate .CSV files until they are merged downstream from a separate application as a final output

Outcomes

  • The machine learning model exceeds the 50% accuracy requirement
    • Vehicle Detection Model Accuracy: 83.9%
    • License Detection Model Accuracy: 98.46%
    • Character Recognition Model Accuracy: 97.49%
  • The system is independent of any internet connection, and does all of the inference on-board and in real-time