{"id":1489,"date":"2021-08-13T10:42:19","date_gmt":"2021-08-13T01:42:19","guid":{"rendered":"https:\/\/blog.testworks.co.kr\/en\/?p=1489"},"modified":"2021-08-24T19:13:55","modified_gmt":"2021-08-24T10:13:55","slug":"deep-learning-for-traffic-counting","status":"publish","type":"post","link":"https:\/\/blog.aiworkx.ai\/en\/deep-learning-for-traffic-counting\/","title":{"rendered":"Deep Learning for Traffic Counting"},"content":{"rendered":"\n

<\/p>\n\n\n\n

\"\"<\/figure><\/div>\n\n\n\n

<\/p>\n\n\n\n

\n
<\/div>\n<\/div>\n\n\n\n
\n
<\/div>\n<\/div>\n\n\n\n
\"\"<\/figure>\n\n\n\n

Changsin Lee l Tech Evangelist l Testworks<\/strong><\/p>\n\n\n\n

<\/p>\n\n\n\n

\n
<\/div>\n<\/div>\n\n\n\n
\n
<\/div>\n<\/div>\n\n\n\n

<\/p>\n\n\n\n

Did you know that traffic counting is behind the rise of Microsoft and thus modern personal computing? Traffic counting is used by city officials to determine transportation needs. The most popular method is by laying a strip of pneumatic rubber tube on the roadway of interest and analyze metrics like vehicle volume, vehicle types, axle counts, etc. This method, however, is not quite accurate or safe so a better method is researched. In this article, you will find how you can build your own traffic counter using a Deep Learning approach (YOLO V5) using a public dataset from the ground up. The entire source code is available in the\u00a0Colab notebook<\/a>. The sample training dataset used is AI Hub dataset number\u00a030743<\/a> which Testworks<\/a> delivered as a public dataset.<\/p>\n\n\n\n

<\/p>\n\n\n\n

\"\uc774<\/figure><\/div>\n\n\n\n

<\/p>\n\n\n\n

\n
<\/div>\n<\/div>\n\n\n\n

<\/p>\n\n\n\n

Problems<\/strong><\/p>\n\n\n\n

Interestingly, before they founded Microsoft, Bill Gates and Paul Allen first tried their hands on automating some manual processes in punching in the pneumatic road tube data. The business partnership was named as\u00a0Traf-O-Data<\/a>\u00a0which later became Microsoft. While the project itself was not very successful on its own, the experience they gained eventually led to founding Microsoft and the rest is history.<\/p>\n\n\n\n

Though pneumatic rubber tube counters are easy and widely used solutions, there are a few problems:<\/p>\n\n\n\n

  1. Setup risks: To set up the tubes, someone has to go out in the middle of a high-traffic roadway. This is risky and poses an insignificant amount of liability issues.<\/li>
  2. Wear-and-tear: Rubber tubes invariably suffer from natural wear-and-tear and thus need to be replaced on regular basis. A more sinister problem is that the device might start reporting inaccurate metrics without people realizing it.<\/li>
  3. Inaccuracy: A bigger problem is that the metrics that can be gathered from pneumatic tubes are rough guestimates at best. Vehicle speed and traffic conditions can all contribute to inaccurate measurement.<\/li>
  4. Scalability issues: While the tube provides a simple method of counting, if you want to gather other information like pedestrians, cyclists, motorists, models of vehicles, etc., there is simply no way the pneumatic rubber tubes can scale up.<\/li><\/ol>\n\n\n\n

    For these reasons, a safer and more scalable method is sought after in the field of Computer Vision, especially Deep Learning. The key to a good performing Deep Learning model is, of course, quality training data. The good news is that there are many publicly available free datasets. In what follows, I am going to show how you can enhance the performance of a traffic counting model through transfer learning on public data. For the experiment,\u00a0a sample surveillance camera dataset<\/a>\u00a0from\u00a0AI Hub Korea<\/a>. There are 100 images taken from a fixed point surveillance camera and they were tagged with bounding boxes.<\/p>\n\n\n\n

    <\/p>\n\n\n\n

    \n
    <\/div>\n<\/div>\n\n\n\n

    <\/p>\n\n\n\n

    Baseline Model: YOLO V5<\/strong><\/p>\n\n\n\n

    YOLO<\/a>\u00a0(You Only Look Once) is an open-source Deep Learning system that is used successfully used for object detection. There are also pre-trained models from which you can train with your own dataset. This is called \u2018transfer learning<\/a>\u2019 and a good way to bootstrap your own vehicle detection system. To do so, let\u2019s get YOLO V5 together with the pre-trained model.<\/p>\n\n\n\n

    \"\"<\/figure><\/div>\n\n\n\n

    <\/p>\n\n\n\n

    Now get some test image data and run through the model. You are welcome to use any images but since we are interested in traffic data, let\u2019s grab some images from a highway surveillance camera\u00a0here<\/a>\u00a0and then run the detect command to obtain yolov5 predictions.<\/p>\n\n\n\n

    \"\"<\/figure><\/div>\n\n\n\n

    <\/p>\n\n\n\n

    Detailed descriptions of what each command line parameter means can be found in the yolov5 repository. For our purposes, we just need to specify the weights file (weights parameter), image size (img parameter), confidence threshold (conf parameter: 0.5 means that the confidence probability should be at least 50%), and the source folder where the test images are located (source parameter). Here is sample output from the detection:<\/p>\n\n\n\n

    <\/p>\n\n\n\n

    \"\"
    YOLO V5 detection result<\/figcaption><\/figure><\/div>\n\n\n\n

    <\/p>\n\n\n\n

    While the pre-trained model did a pretty good job in detecting and recognizing vehicles, it missed a lot of vehicles in the back. Even worse, in some cases, it recognized a row of cars as a train as in the below example. Ouch!<\/p>\n\n\n\n

    \"\"<\/figure><\/div>\n\n\n\n

    <\/p>\n\n\n\n

    \n
    <\/div>\n<\/div>\n\n\n\n

    <\/p>\n\n\n\n

    Train a Vehicle Counting Model<\/strong><\/p>\n\n\n\n

    To enhance the pre-trained model, we need some labeled data. AI Hub is a public dataset repository in Korea and we can find a\u00a0labeled highway image dataset<\/a>\u00a0from a surveillance camera. The whole dataset contains 500,000 images. For an illustration purpose, let\u2019s just use the sample dataset that has 100 images labels. The same sample test image shows the target labels that YOLO V5 missed:<\/p>\n\n\n\n

    <\/p>\n\n\n\n

    \"\"
    Trains in a highway? No way!<\/strong><\/figcaption><\/figure><\/div>\n\n\n\n

    <\/p>\n\n\n\n

    To train the YOLO model, we need to configure a few things. The most important thing is the label format and the train configuration file. Let\u2019s look at the configuration file first. This file must contain the datasets you want to use (train, validation, and test data folders), the number of classes, and the class names. Since our purpose is vehicle counting, there are only three vehicle classes. Here is the entirety of the train configuration YAML file.<\/p>\n\n\n\n

    \"\"<\/figure><\/div>\n\n\n\n

    <\/p>\n\n\n\n

    Second, for each image file, there must be the corresponding *.txt file with the same file name. In each label text file, the labels must be of the following format:<\/p>\n\n\n\n

    \"\"<\/figure><\/div>\n\n\n\n