Video Analysis of Object Tracking and Detection on Drone Captured Video

Abstract

Object detection, tracking, and distance measurement play a crucial role in visual surveillance systems, particularly in aerial scenarios. The VisDrone dataset, a benchmark dataset for aerial computer vision, provides a challenging testbed for developing robust algorithms in these domains. In this study, we propose a comprehensive framework for object detection, tracking, and distance measurement using the VisDrone dataset. First, we employ a state-of-the-art object detection model based on deep learning, such as YOLOv7, to detect objects of interest in aerial images. The model is trained on the VisDrone dataset, which consists of various object categories, including pedestrians, vehicles, and cyclists. Next, we incorporate object tracking algorithms to track the detected objects across consecutive frames. Robust tracking algorithms, such as Kalman filtering, particle filtering, or deep learning-based trackers, are utilized to handle occlusions, scale changes, and abrupt object movements. The tracking algorithm ensures continuous localization and association of the detected objects, enabling their persistent tracking. Here we have used Deep Sort object tracking algorithm. To estimate the distance of the detected and tracked objects from the aerial camera, we utilize geometric and perspective cues. By leveraging the known camera parameters and image geometry, we can calculate the distance based on object size, position, and perspective distortion. We compare our results with existing approaches and demonstrate the effectiveness of our approach in challenging aerial scenarios. The proposed framework has wide-ranging applications in aerial surveillance, traffic monitoring, crowd analysis, and urban planning. It enables real-time object detection, tracking, and distance measurement from aerial imagery, providing valuable insights for situational awareness and decision-making processes in various domains.