REAL-TIME MULTI-THREADED BASED OBJECT DETECTION AND COUNTING SYSTEM USING YOLOV8 AND BYTETRACK

Abstract

Object detection from real-time video streaming is a challenging task, and this paper presents an innovative approach for a real-time object detection and counting system with YOLOv8. For the multi-object tracking from the detected objects of YOLOv8, ByteTrack is used. The designed system is used for long-duration video streams of about two hours of continuous video segments extracted from the static surveillance feeds. The multithreaded architecture utilized here is to read the frame with maximized throughput, with mitigated frame loss. Meanwhile, the object counting is performed with the user-defined virtual line with reduced jitter using the bounding box smoothing. The duplication of object counting is avoided with the inclusion of an intelligent ID system. The object bounding boxes, live count overlay, and center points are added to the disk in the real-time processing for the robust video writing approaches. To avoid resolution problems, the custom re-encoded pipeline is used. This can be used in real-time applications including industrial automation, retail analytics, and traffic monitoring.