Real-Time Multi-Viewpoint Object Detection for High-Speed Omnidirectional Scanning Shooting System

Abstract

Obtaining object-detection results through multi-viewpoint systems is a promising approach for gathering the surrounding information for mobile robots. In this study, we realized a low delay real-time multi-viewpoint object detection for an omnidirectional scanning shooting system by proposing a dual-stage detection approach with intermediate frames. This approach can synchronously search for valid viewpoints containing the desired objects, and perform low-update-delay real-time object detection on selected viewpoints with intermediate frames. We generated panoramic detection frames as real-time system output by projecting the individual viewpoint detection results. In an outdoor experimental scenario with 10 fps omnidirectional scanning shooting accommodating up to 40 viewpoints, our approach generated 20 fps real-time panoramic detection frames presenting the movement of the observed targets.