In this project we attempt to use a YOLOv5 model to detect weapons in various scenarios.
We use Transfer-Learning by performing additional training on a pre-trained network, using the new dataset, to achieve this task.
This is a class project as part of EE046211 - Deep Learning course @ Technion.
Itamar Ginsberg • Alon Nemirovsky
To ensure citizens' safety, we want to enable security cameras to automatically detect a person carrying a weapon, so security services can be quickly alerted and can then decide on the best course of action, thus dramatically reducing response time to save lives. For this purpose, we take advantage of an advanced neural-network architecture utilized for detection tasks in other domains through the technique of transfer-learning and fine-tuning.
We use the ari-dasci/OD-WeaponDetection dataset, mainly 'Knife Detection' and 'Pistol Detection'.
The data includes 3000 pistol images and 2078 knife images, many of which feature life-like scenarios.
Image examples from the dataset:
Annotation examples from the dataset: (matching the images above)
[0 0.378 0.387 0.205 0.332]
[0 0.531 0.536 0.179 0.326]
[0 0.501 0.506 0.424 0.364]
[1 0.351 0.669 0.163 0.105]
[1 0.521 0.449 0.647 0.601]
[1 0.618 0.297 0.116 0.081]
Each line in a file includes the class number of the detected object (0 - pistol, 1 - knife) and 4 points (normalized) indicating the object's bounding box. There can be more than one line in an annotation file, indicating more than one object in the scene.
YOLO (You Only Look Once) divides an image into a grid system, where each grid detects objects within itself. It is a very fast algorithm that can be used for real-time object detection based on data streams, while requiring very few computational resources.
The architecture of a detection network is generally made up of a backbone (responsible for feature extraction) and a head (responsible for bounding-box and label prediction). In many instances, the model also includes a 'neck' between the backbone and head (responsible for feature fusion).
For YOLOv5, the backbone is carried out by the CSPDarknet53 CNN, the neck uses PANet and the head is a YOLO layer.
While yolov4(April 2020) introduced new augmentations such as Mosaic and Self-Adversarial-Training, YOLOv5, released a month later by Glenn Jocher, features implementation based on the PyTorch framework and the use of .yaml files for configuration.
Currently (1/22) there is no official paper on this model.
Ultralytics offers a choice of 5 YOLOv5 models, differing mainly in size and complexity:
| File name | Description |
|---|---|
DL_project_code |
Main code file - includes setup, data reformatting and reorganization, training the model and using the outcome to detect new images |
yolov5/data/Weapons.yaml |
Definition of dataset location, number of classes and names |
yolov5/data/hyps/hyp.scratch.yaml |
Definition of the model's hyperparameters and augmentation parameters |
yolov5/models/yolo5*.yaml |
Configuration file for { Nano (n) / Small (s) / Medium (m) / Large (l) / XLarge (x) } models |
yolov5/runs |
Location of runs' data and results for both training and detection |
repository_images |
Images used for preview in README.md file |
As shown in the resulting images above, our model was able to detect the desired objects, even in dense or very small instances, both of which are known to be weaknesses of the classic YOLO algorithm.
[1] ari-dasci/OD-WeaponDetection full dataset
[2] v7labs - YOLO: Real-Time Object Detection Explained
[3] Ultralytics YOLOv5 repository
[4] Ultralytics YOLOv5 documentation
[5] Bochkovskiy, A., Wang, C.Y. and Liao, H.Y.M., 2020. Yolov4: Optimal speed and accuracy of object detection (arXiv preprint arXiv:2004.10934.)
[6] analytics-vidhya (Medium): Object Detection Algorithm — YOLO v5 Architecture