Towards Condition-Robust Palm Vein Recognition: Dataset and Performance Analysis

Palm vein biometrics is contactless identification through vascular vein patterns. The paper presents a new dataset of 500 palm vein images from 100 individuals in 5 conditions (normal, hot, cold, dusty, and lotion-applied). In contrast to current benchmarks, the dataset directly simulates environmental and physiological variations. It compares three feature-extraction pipelines (Kumar Gabor, IUWT-SAD and Maximum Curvature) to a proposed multi-feature ensemble SVM. The proposed SVM uses HOG, LBP and Gabor features. In all cases, the ensemble achieves a mean EER of 4.0 %, TAR FAR=10-3 =72.3%, and AUC= 0.963, which is on par with Kumar Gabor (EER 8.8%), MC (EER 13.1%), and IUWT-SAD (EER 16.4%). Performance is consistent in response to temperature changes. There is only slight performance deterioration in the presence of surface contaminants (dust, lotion). Calibration analysis indicates low error (ECE < 0.02, Brier < 0.03). A throughput of up to 12 images per second is achieved with the proposed feature pipeline. The results demonstrate that ensemble fusion is highly effective for condition-resilient palm vein recognition. The new dataset offers a good reference point for estimating real-world resilience beyond laboratory tests.