Inspection and Counting of Spur Gear Teeth Defect using YOLOv7 and DeepSort

Overview

This implementation is the inspection of spur gear teeth defects using state-of-the-art YOLOv7 Object Detection and DeepSort Object Tracking Algorithm. The model was trained on custom collected dataset which are images of 3D printed gears, and the model was trained to achieve over 90% mAP score as application area is intended for product quality inspection. The detection result of the model is then itegrated into DeepSort tracking algorithm to retrieve the track ID of detected objects. The track ID of specific detected label is stored into data structure when the object reach within ROI area, and finally the counts of defect, good and total products are measure based on the numbers of IDs stored in the data structure. The general pipline of the system is as shown in the following figure.

Detection + Tracking (Counting)

---

Contents

• Inference Speed of Detection and Detection plus Tracking
• Dataset Preparation
• Installation
• Run Demo
• Model Training
• Model Evaluation
• References

---

Inference Speed of Detection and Detection plus Tracking

System Specification

• OS : Ubuntu 22.04 LTS
• CPU : Intel Core i5 6200U 2.30GHz * 4
• RAM : DDR4, 4GB
• GPU : NVIDIA GeForce 920Mx, 2GB
• GPU Drvier & CUDA Version : 515.105.01 | 11.7

Inferencing	FPS on above System Spec
YOLOv7 (Detection)	388.6ms
YOLOv7 + Deepsort (Detection+Tracking)	350.3ms

---

Dataset Preparation

The dataset that is being trained on is custom dataset. You can get gear images here-> (link soon), annotate with labelImg as the steps below.

Step-1: Data Annotation

• Installing labelImg Data Annotation Tool;

git clone https://github.com/heartexlabs/labelImg.git   # first clone the original Repo on specific directory

conda create -n LabelImg python=3.9.13

conda activate LabelImg
pip install labelimg

• Creating pre-defined classes.txt file in images(dataset) directory with our interest class-names;

Good Teeth
Defect Teeth

• Starting Data Annotation for YOLOv7;

cd [the cloned LabelImg Repsitory]

python labelImg.py [/path/to/images-datasets] [/path/to/classes.txt]

Make sure YOLO type > Create RectBox > Save > Next

Step-2: Data Splitting

python split_dataset.py --folder [annotated-dataset-folder-name] --train [80] --validation [10] --test [10] --dest [/desinated-path-name-dataset]

Step-3: Making configuration file

• The parameters in gear_detect.yaml are needed to change.
  • add the training, validation and testing data path to train, val, and test respectively
  • add the numbers of class to class to nc, and class names to names parameters
• The configuration parameter nc number of class in yolov7-gear-detect.yaml is also needed to changed, and we chage to 2 for this case.

---

Installation

Step-1: Install NVIDIA GPU driver & CUDA if it's not already installed on the machine (in Ubuntu)

# Installing the driver with your desired verion number
	sudo apt install libnvidia-common-515
	sudo apt-get -y install libnvidia-gl-515
	sudo apt install nvidia-driver-515-server
	
# After installing with all above commands, reboot ubuntu
	sudo reboot
	
# After reboot, check the smi table
	nvidia-smi	

Step-2: Creating environment & install dependencies

• Install anaconda, create conda environment

conda create -n spur-gear-teeth-defects python=3.10.6 

• Install PyTorch & required packages

conda activate spur-gear-teeth-defects

pip3 install torch torchvision torchaudio  # Recommend: Go to pytorch.org and get the command according to desirements

pip install -r requirements_gpt.txt

For training YOLOv7 object detection model and running demo, we need to install YOLOv7 Weights in the the working directory. If you want to train with other YOLOv7 weights, you can check here. To do inferencing with my trained model, download it from here, and paste it to the directory runs/train/yolov7-defect-detect-three2/weights/

---

Run Demo

For running the demo of the project using the trained model, three main cases of demo can be ran including normal defect detection, detection and tracking of the product with counting inspection results, and the results can be saved as CSV file for visualizing inspection charts in Excel. The demo result files are saved in runs/detect/ path after running each python script.

Run for detection

To run detection of spur gear teeth defects, we need to run detect.py python script.

• Argument lists;
  • --weights - path to the trained model, runs/train/.../last.pt
  • --source - path to inference data source
  • --view-img - display detecting frame

python detect.py --weights runs/train/yolov7-defect-detect-three2/weights/last.pt --source inference/videos/video_6.mp4 --view-img

Run for detection + tracking -> Counting

To run counting the inspecting result of gear product through detection and tracking, we need to run the implemented detect_track_count.py python script, and the argument are the same as detection.

python detect_track_count.py --weights runs/train/yolov7-defect-detect-three2/weights/last.pt --source inference/videos/video_4.mp4 --view-img

Run for saving inspected data as CSV file

For saving the data of inspecting result as CSV file, --save-csv is needed to pasre in the command. This saved file is going to save in ./runs/csv/ path, and this csv file is used to visulize the qulity control plot and charts as following figure.

python detect_track_count.py --weights runs/train/yolov7-defect-detect-three2/weights/last.pt --source inference/videos/video_4.mp4 --view-img --save-csv

Model Training

To train the model on your machine, make sure YOLOv7 weight file haved already downloaded.

python train.py --workers 0 --batch-size 1 --device 0 --data data/gear_detect.yaml --img 640 640 --cfg cfg/training/yolov7-gear-detect.yaml --weights yolov7.pt --name yolov7-defect-detect-three --hyp data/hyp.scratch.custom.yaml --epochs 200

---

Model Evaluation

To evaluate the run model on Tensorboard, run the following;

tensorboard --logdir runs/train

---

References