Splash Detection Research Of UWA

This project aimed to use yolo v4 to detect the splash of Western Australia's high speed road.

Implementation of YOLO V4
The implement of yolo v4 comes from Alexeyab's Github:https://github.com/alexeyab/darknet, which provides the source code of as well as darknet for cross OS platform. You need to first compile then run the demo.

The introduction covers most your demand of the task. It provides step by step tutorial of:

• How to compile it

• How to run detection with pre-trained weights of image and vedio

• How to train it for custom object detection ·How to test the validation

• How to write your confiuration file

• How to adjust parameters

Environment
Th environment I used is Google Colab, with tutorial.
After you learned colab, you can apply this projet on it follw this tutorial

Step by step guide

1. Firstly, you can access My Google Drive¹
2. Move the folder 'Yolo' into the same dirctory in your Google Drive for Colab's run.
3. If you don't do step 3, you need to replace all the path in below's file.
4. Open my Colab file and run it cell by cell to detect the the splash for a vedio
5. The final Command is for train the model with the splash dataset.

Files Introduction
There are only 3 files need you to write for train the network.

1. obj.Data -- which indicates your trainning set and valid
2. obj.cfg -- which contains the configuration of your custom network
3. obj.names -- which contains the name of your aimed object. Besids, here are some other files:

And also other files as blow. Those file with link means you can't find it in github since the size exceed 100M, you can only get it from Google Drive:

File name	Details
Yolo-obj_final.weights	The weights I trained to detect splash. You can directly apply it to replace the pretrained-weights of Alexey's project to get the detector.
OriginalVedio	It provides a original vedio without result
ResultVedio	It provides a vedio with a result of my detection
Train.txt	The trainning set index of images.
Valid.txt	The valid set index of images.
obj folder	All the images and labels file.

Command
Here is the command to use the yolo network:

Train : darknet detector train <.data> <.cfg> <PreTrainWeights> -dont_show
Train by mAP ： darknet detector train <.data> <.cfg> <PreTrainWeights> -map
Detect Image ： darknet detector test <.data> <.cfg> <SelfTraiinedWeights>
Detect Vedion ： darknet detector demo <.data> <.cfg> <SelfTrainedWeights> -dont_show <InputVedioPath> -i 0 -out_filename <OutputVedioPath>

Note

1. When follow Alexey's guide to detect vedio, do remember the path of your vedio should be relative path from darknet. Abosolute path may cause error.

---

File Path

All the files you need to change the path to run.
In the case to run on Google Colab, you first need to mount your Google Drive to Colab.
Then, you will find your Google Drive at "/content/gdrive/My Drive/"
Then I created a folder named Yolo to store Alexey's repository darknet.
Here listed the path of all my files, if you want to use my cfg, make sure your files are in the same location.

configuration files:

• cfg: "/content/gdrive/My Drive/YOLO/darknet/cfg/yolo-obj.cfg"
• data: "/content/gdrive/My Drive/YOLO/darknet/data/obj.data"
• Self-trained Weights: "/content/gdrive/My Drive/YOLO/darknet/backup/yolo-obj_final.weights"
• names: “/content/gdrive/My Drive/YOLO/darknet/data/obj.names”
• PreTraind Weights: "/content/gdrive/My Drive/YOLO/darknet/yolo4.conv.137”

data files:

data	path
Images	"/content/gdrive/My Drive/YOLO/darknet/data/obj/<ImageName>.jpeg"
Labels	"/content/gdrive/My Drive/YOLO/darknet/data/obj/<labels>.txt"
Train index	"/content/gdrive/My Drive/YOLO/darknet/data/<train>.txt”
Valid index	"/content/gdrive/My Drive/YOLO/darknet/data/<valid>.txt”

Note

1. When use the path in command line, do use "path" instead of path, because there is a blank of My Drive which may lead to error with out ""
2. When follow Alexey's guide to detect vedio, do remember the path of your vedio should be relative path from darknet. Abosolute path may cause error.

Footnotes

1. Here is almost the same as github only added some config flie to do splash detection ↩