Violence Detection Using YOLOv8 - Towards Automated Video Surveillance and Public Safety

Title:

• Violence Detection Using YOLOv8: Towards Automated Video Surveillance and Public Safety

Team-mates:

• Md. Alahi Almin Tansen
• Umayer Mohammad Affan
• Afruja Sultana Muniya
• Md. Delower Hosen

Supervisor:

• Sovon Chakraborty
• Md. Obaidur Rahaman

Selected Dataset:

• Custom training dataset : Roboflow Dataset
  • Total = 2834 images
    • Train = 1969 images
    • Valid = 575 images
    • Test = 290 images
• Video dataset: Kaggle Dataset (Not using this as it is same dataset as our selected image dataset)
  • Total = 2000 videos
    • Non-violence = 1000 videos
    • Violence = 1000 videos
• Video dataset: RWF-2000: An Open Large Scale Video Database for Violence Detection
  • Total = 2000 mixed videos

Selected model

• Yolov8
• Some CNN models
• Yolo-NAS

Update 01

• Model used YOLOv8s
• Number of epochs = 25
• Batch size = 16
• Total training time = 0.255 hours
• Confidence threshold = 0.25
• Prediction on videos = 10 videos

Update 02

As advised by the supervisor we used some CNN models and Yolo-NAS model and compare each of those models.

CNN models we used:

• VGG16
• VGG19
• ResNet152V2
• InceptionV3
• MobileNetV2
• DenseNet201

CNN models:

• Number of epochs = 25
• Batch size = 32
• Loss function used = smooth_l1_loss
• Intersection Over Union (IOU) is observed in train , test , validation
• Total training time:
  • VGG16 - 1640.69 seconds
  • VGG19 - 1962.71 seconds
  • InceptionV3 - 1204.79 seconds
  • MobileNetV2 - 1023.70 seconds
  • DenseNet201 - 1547.24 seconds

Yolo-NAS model:

• Model used yolo_nas_s
• Number of epochs = 25
• Batch size = 16
• Caching annotation time (minutes) = Train dataset-07:35 Valid dataset-02:10 Test dataset-01:03
• Total training time (minutes) = 75.4 minutes
• Confidence threshold = 0.25
• Prediction on videos = 10 videos

Update 03

Vehicle detection and count

Selected Dataset:

• Custom training dataset : Roboflow Dataset
  • Total = 17491 images
    • Train = 15296 images
    • Valid = 1458 images
    • Test = 737 images
• Model used YOLOv8s
• Number of epochs = 15
• Batch size = 16
• Total training time = 2.568 hours
• Confidence threshold = 0.25
• Prediction on images = 737 images
• For object tracking = ByteTrack
• For Line drawing ,annotation , coloring frame by frame = Supervision
• Prediction and tracking on videos = 2 Videos

Update 04

Sample images from the dataset & few changes.

• Viewed some of the sample images from the dataset to include in paper.
• Kaggle video dataset won’t be used in model testing
• RWF-2000 video dataset will be used in model testing