A FINE-GRAINED OBJECT DETECTION MODEL FOR AERIAL IMAGE BASED ON YOLOv5 DEEP NEURAL NETWORK

Abstract

A real-time visual tracking system with an active camera is implemented and described in the study. The system's purpose is to track human mobility indoors. Frame differencing and camera motion compensation are the foundations of quick and easy motion detection techniques that enable real-time tracking. The outcomes of the online person tracking are shown. The multiple objects tracking approach maintains a graph structure where it maintains multiply hypotheses on the number of aids the object trajectories in the video have, based on initial object detection results in each image, which may have missing and/or inaccurate detection.The graph is extended and pruned based on the image information, which also establishes the optimal hypothesis. Although the tracking process confirms and validates the detection over time, image-hued object detection makes a local judgment. As a result, it can be seen of as temporal detection, which makes a global decision over time. The multiple object detection approach provides the object detection module with feedback in the form of object position predictions. Thus, object detection and hauling are strongly integrated in the technique. Multiple object tracking results are provided by the most plausible hypothesis. Results from the experiment are shown.