Multi-view Multi-camera Object Detection and Tracking: A YOLOv7 and DeepSORT-Based Approach

Abstract

Multiview object detection and tracking refer to the process of detecting and tracking objects from multiple viewpoints or perspectives, often using multiple sensors or cameras. This work investigates the efficacy of integrating a cutting-edge object detection model with a robust multi-object tracking algorithm, for multi-view multi-camera object detection and tracking (MVMCT). We employ a late fusion method, each camera image undergoes independent processing by a state of the art object detection model YOLO to generate detections. Then we employ a robust multi-object tracking algorithm DeepSORT that handles the association of these detections across cameras and manages tracks over time, utilizing Kalman Filters and appearance modelling. The study showcases the real-time object detection capabilities of YOLOv7 in MVMCT scenarios and assesses DeepSORT's performance in associating detections and sustaining tracks across multiple views. Comparative analyses against other methods in multi-camera object detection and tracking (MCODT), considering various conditions such as dynamic environments and occlusions, are conducted. The integration of YOLOv7 and DeepSORT demonstrates notable accuracy and resilience in multi-view object detection and tracking tasks. The late fusion approach offers adaptability and modular integration with diverse object detectors. The system exhibits promising outcomes in challenging scenarios, underscoring its potential for practical applications. This research contributes to advancing the field of MCODT by presenting an efficient and effective solution for precise object localization and tracking across multiple cameras, with potential applications in surveillance, traffic monitoring, and autonomous vehicles. The evaluation was performed on the Multi-View Multi-Camera dataset from EPFL CVLAB.