Advanced Video Analytics: MediaPipe, Bounding Boxes, and Graph-Based Tracking for Object Detection

Detecting and tracking(D &T) moving objects presents important preliminary steps in the processes of object recognition, contextual analysis, and indexing within visual surveillance systems. D &T refer to the procedure of identifying and monitoring the presence and motion of individuals within a specified area, often using employing machine learning techniques. Multiple methods exist to represent object detection and tracking, including the use of graphs. Graphs offer a natural means to depict the movement of objects. However, creating a model capable of operating within such a structured domain is an impressive task. In this paper, we introduce an innovative model for detecting and tracking objects in videos based on Graph Neural Networks (GNN). This model converts video into a graph, where nodes correspond to coordinates extracted from bounding boxes detected using MediaPipe, while edges represent connections between objects on consecutive frames. These connections are established based on the Euclidean distances between the centers of the objects. Ultimately, we obtain a graph that undergoes classification through both edge and node classification methods. During the evaluation phase, our approach was rigorously tested on the MOT20 dataset, employing a range of metrics, such as Multiple Object Tracking Accuracy (MOTA), Identity F1 (IDF1), and Mostly Tracked (MT). This thorough assessment not only demonstrated significant advances over state-of-the-art techniques.