Object detection for harvesting robots in low-light orchards via YOLO and self-calibrated illumination

Abstract

The main reason for a high miss rate in robotic agricultural harvesting is still the target feature degradation caused by unstructured canopies and low light. SCI-YOLO is a lightweight visual perception network in this paper that can perform detection accurately in various light conditions through self-calibration. Given that the data at night is relatively expensive to obtain, a Gamma-correction gradient simulation will be employed instead. The core network adds a Self-Calibrated Illumination (SCI) module to the YOLO base and uses a brightness-aware dynamic routing mechanism for adaptive feature fusion. Since SCI only needs one basic block during inference, it recovers the obscured spatial texture before the main feature extraction with very low operational latency and no image artifacts. According to the test results, the SCI-YOLO framework has achieved an improved mAP and strictly increased the maximum recall (Rmax) for shadowed objects by 20.0% in very dark conditions compared to the original baseline.