Analysis of Live Video Object Detection using YOLOv5 and YOLOv7
Thammi Reddy Konala, Anusha Nammi, Divya Sree Tella
Abstract
With wide use of IoT devices across all the applications has generated huge amount of image data. With the envision of smart city project in India has provisioned for collecting collating and analysing of traffic video stream data for the smooth flow of traffic. In the process of decision making how effective to identify the identity of an offender in traffic violation. The paper is mainly focused on evaluating two most prominent object detection methods i.e. YOLO (You Only Look Once). For training these models, we have considered common Objects in Context (COCO) dataset. The YOLOv5 make use of anchor boxes to detect the objects in the image and it is successful in prediction based on probabilities.The YOLOv5 has some limitation like misclassification when the objects are small in size, fails in angle deflection and has a limited scalability. The YOLOv7 is a real time object detection model uses instance segmentation. The implemented model of YOLOv7 with a live video to detect whether a two wheeler rider is wearing a helmet or not, as wearing a helmet is one of the safety measure while driving on busy roads in India. The YOLOv7 has achieved higher mean average accuracy than YOLOv5.