Multi-Sensor Small Target Detection for UAV Platform

Abstract

To break through the performance limit of single-source sensing in a dynamic environment and solve the reliability bottleneck of low-and-slow small target detection, this paper systematically combs the technical basis of four types of modes: visible light, infrared, millimeter-wave radar, and event camera. Visible light imaging has high spatial resolution and semantic expression ability, but it depends heavily on lighting conditions. Infrared can realize passive all-weather sensing, which is limited by low resolution and thermal noise. Millimeter wave radar has strong penetration and accurate range and velocity measurement, but the lack of angular resolution restricts fine-grained classification. The event camera suppresses motion blur with a microsecond response and an ultra-wide dynamic range, but significantly attenuates the response to static targets. The four are highly complementary in space-time characteristics, which determines that multimodal fusion is the inevitable path to expand the perceptual envelope. However, there are still many problems, such as high cost of cross-modal registration, strong dependence on intensive annotation, degradation of multi-modal synchronization in extreme environments, and limited airborne computing power and communication bandwidth.