The presence of electric bicycles in elevators poses serious safety hazards, necessitating reliable automatic detection solutions. To address these issues, this paper proposes EBike-YOLO, an enhanced real-time detection model based on YOLOv10n, specifically optimized for elevator environments. EBike-YOLO introduces three key innovations: 1) the PIoU2_NWD loss function, combining adaptive penalty mechanisms, anchor-quality-aware gradient adjustments, non-monotonic attention, and Wasserstein distance for significantly improved object localization; 2) the EnhancedPSA structure, refining self-attention and feed-forward network architectures to enhance feature extraction; and 3) the C2f_Calibration module, incorporating self-calibration operations to robustly manage diverse orientations and scales of electric bicycles. Extensive experiments on a custom elevator dataset demonstrate substantial improvements over the YOLOv10n baseline, achieving notable enhancements in precision (4.1%), recall (0.3%), F1 score (2.0%), and mAP@50 (2.23%). Further validations on standard benchmark datasets (VisDrone2019, VOC2007, VOC2012) confirm the model's strong generalization capability, underscoring its applicability beyond specific elevator scenarios. These results clearly highlight the effectiveness, novelty, and practical value of the proposed model for diverse real-world detection tasks.