Robust Out-of-distribution Detection in Orchard Environments via Fusion of Higher-order Statistics

Ensuring the safety of autonomous robots in unstructured environments, such as orchards, is challenging because conventional semantic segmentation models often misclassify out-of-distribution (OoD) objects unexpected obstacles as traversable areas. These OoD objects lead to silent failures that are difficult to detect but critical for safe operation. A widely used post-processing baseline, Standardized Max Logits (SML), recalibrates per-class logit statistics using the mean and variance, but this first- and second-order summary is still insufficient to characterize complex logit distributions in orchard scenes. In this paper, we propose a simple, training-free OoD detection pipeline that augments SML with higher-order statistics skewness and kurtosis computed from per-pixel logit vectors. After min-max normalization, the three scores are fused by element-wise multiplication, which acts as a logical AND gate and raises an alarm only when all indicators consistently suggest an anomaly. Experiments on the large-scale 'Robot Driving in Orchard' dataset (approx. 170,000 images) from AI Hub demonstrate improves all evaluation metrics compared with existing logit-based methods, achieving an Area Under the Receiver Operating Characteristic (AUROC) of 94.57% and an Area Under the Precision-Recall Curve (AUPRC) of 22.35%. In particular, the False Positive Rate (FPR), which is directly related to safety, is reduced by 7.29% relative to SML, significantly minimizing the risk of false alarms for practical implementation [A4] while incurring negligible additional computational cost. These results show that incorporating higher-order logit statistics is an effective way to enhance the robustness of existing segmentation models in real-world agricultural environments and potentially other unstructured domains.