Video Segmentation using YOLOv5 for Surveillance

One of the most major issues associated with computer vision applications includes image and video segmentation. The objective of video segmentation is the extension of the image segmentation task from the image sector to the video sector. The problem is the simultaneous detection and segmentation of objects in image frames of videos. Many new methods typically develop complicated pipelines to handle the task of video segmentation. Previous methods generally first obtain an image frame or clip from the video, do the segmentation task and piece together those incomplete results by tracking or matching them with a very low accuracy of less. This paper proposes a new way of video segmentation using the YOLOv5 algorithm which puts forward the use of a neural network from beginning to end (the complete process from the beginning to the end) that makes predictions of bounding boxes and class probabilities concurrently. When provided with a video clip containing multiple frames of images as input from the user, the YOLOv5 algorithm will output the objects in each frame of the video in the same order as given by the user. Also, the YOLOv5 and the Mask R-CNN [12] algorithms are compared in terms of computational speed and the comparative implementation of both the algorithms show that the YOLOv5 is much faster than the Mask R-CNN algorithm.