Introduction
This project is aimed at building an object detector, using Tensorflow and Python, to find and label highway overpass signs in an image.
Motivation
You may have seen news reports of trucks running into bridges that they do not fit under. There even is a site dedicated to one specific low bridge http://11foot8.com/.
I thought it would be interesting to come up with a solution that could find bridge overpass height signs and identify the height of the bridge. Then, if the vehicle height is known, the system could notify the driver that they are about to crash into a bridge that is too low for the vehicle.
Setup
I ran everything on Windows 10, but the same setup should work on Linux. Download the v1.13 release of the Tensorflow Models directory: https://github.com/tensorflow/models/tree/r1.13.0
Then download this repository into the models/research/object_detection folder.
Installing Tensorflow
I followed this guide for Installing Tensorflow (and for training): https://github.com/EdjeElectronics/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10
I used the following software versions:
| Software | Version |
|---|---|
| Tensorflow | 1.15.0 |
| Python | 3.5.6 |
| CUDA | 10.0 |
| CUDNN | 7.6.5.32 |
| TensorRT | 6.0.1.5 |
Running
Run Overhead_Sign_Detection.py from the models/research/object_detection folder. You can see a sample image input in the main function of the script.
Results
The test images data set worked fairly well at detecting both the overhead signs and the height. The height was determined by detecting each digit individually (0-9) and then merging them. By detecting each digit individually, the detector could determine the height of an arbitrary sign, rather than only heights with training data.
In the videos the sign detector worked well, but not the digit/height detection. I took videos with my cellphone, where as the training/testing images were scrapped from google maps. I cropped the images to reduce the size (~250x200 pixels) and to speed up training time. However, the videos from my cellphone were 1920x1080p. The detector didn't work on frames so large. I had to scale down the videos to 426x240p. In the process the images lose a lot of detail and the detector did not work as well. You can see this in the video at the top.
