rumeysakeskin / yolo-darknet-video-and-image-detection-traffic-signs Goto Github PK

In this project, German Traffic Sign Detection Benchmark Dataset was used. Dataset includes the 900 training images (1360 x 800 pixels) in PPM format, the image sections containing only the traffic signs, a file in CSV format with the ground truth, and a ReadMe.txt with more details.

Note: Do not need to the download dataset, we will download it in training_YOLOv3_Darknet.ipynbb 🥳

Open training_YOLOv3_Darknet.ipynb in Colab and run the code.

• In the YOLO format, every image in the dataset has a single text file. If an image has no objects there is no text file for that image.
• Inside the text file, each row contains the following information: The first element of each row is a class id, then bounding box properties (x, y, width, height). Bounding box properties must be normalized (0–1). (x, y) should be the mid-points of a box.

images
→ 0001.jpg
labels
→ 0001.txt
 ├── class_id, x_centre, y_centre, width, height

• Also we need to create the following files before training:

  ├── train.txt
  ├── test.txt
  ├── ts_data.data
  ├── classes.names

These files' formats are as following:

train.txt and test.txt

/content/FullIJCNN2013/00539.jpg
/content/FullIJCNN2013/00207.jpg
/content/FullIJCNN2013/00075.jpg
.
.

classes.names

prohibitory
danger
mandatory
other

ts_data.data

classes = 4
train = /content/FullIJCNN2013/train.txt
valid = /content/FullIJCNN2013/test.txt
names = /content/FullIJCNN2013/classes.names
backup = backup

yolov3.cfg includes parameters for training, structure of CNN layers, last three YOLO layers.

filters = (classes + coordinates + 1) * masks

max_batches = classes * 2000 (not less than 4000)

steps = 80% and 90% of max batches (max batches represent total number of iterations for training)

For traffic sign dataset:

filters = (4 + 5) * 3 = 27
max_batches = 4 * 2000 = 8000
steps = 6400, 7200
classes = 4
batch = 32 (training), 1 (testing)
subdivisions = 16 (training), 1 (testing) - (represent number of minibatches in one batch)

• Prepare yolov3_train.cfg and yolov3_test.cfg files according to your custom dataset following above instructions and put under the darknet/cfg/ location.

After preparation of all files we will train the the custom dataset with following command:

!./darknet detector train cfg/ts_data.data cfg/yolov3_train.cfg darknet53.conv.74 -dont_show

Weights will be saved every 100 iterations in the following path. Now, we will define the best one to use for detection making sure that there is no overfitting.

├── darknet/
    ├── backup/
        ├── yolov3_train_1000.weights
        ├──   ...
        ├── yolov3_train_last.weights
        ├── yolov3_train_final.weights

There is special command in Darknet framework that calculates mAP.

!./darknet detector map cfg/ts_data.data cfg/yolov3_ts_train.cfg backup/yolov3_train_1000.weights

Continue checking for all weights in order to define the weights with biggest mAP.

• Put traffic-sign-test.mp4 under the darknet/data/, and yolov3_train_1000.weights to darknet/weights/ location.

!./darknet detector demo cfg/ts_data.data cfg/yolov3_ts_test.cfg weights/yolov3_train_1000.weights data/traffic-sign-test.mp4 -out_filename traffic-sign-to-test.avi -dont_show

Open the traffic_sign_image_detection.ipynb notebook and run the script for your test image.

Open the traffic_sign_detection_on_video.ipynb notebook and run the script for your test video.