Optimized YOLO Architecture Detects Electric Bicycles in Crowded Elevators

A research team has introduced an ultra-lightweight object detection model named VGWS-YOLO, designed specifically to detect electric bicycles entering high-rise residential elevators and help mitigate lithium-ion thermal runaway fire risks. When electric bicycle batteries undergo thermal runaway in the narrow, enclosed spaces of elevator cabins, rapid deflagration can generate high-temperature toxic fumes and heavy smoke, endangering passengers, disabling elevator mechanics, and blocking evacuation routes. Traditional community oversight methods, such as manual inspection and physical barriers, frequently leave supervision gaps and fail to sustain continuous operational management. To counter these vulnerabilities, automated surveillance based on edge computer vision has emerged as an essential intervention, although real-world deployment on resource-constrained embedded sensing hardware has historically been compromised by dense target occlusions, optical reflections, and severe computational limits.

Early visual recognition systems in residential buildings depended heavily on handcrafted feature descriptors, such as color distributions, target contours, and Histograms of Oriented Gradients, paired with conventional classifiers like Support Vector Machines. While these classical machine learning pipelines required minimal computational power, they lacked deep representational capacity and suffered catastrophic performance drops when confronted with lighting fluctuations, severe passenger occlusions, and shifting perspective angles. Modern deep learning architectures, particularly the You Only Look Once series, eliminated manual feature engineering by learning hierarchical visual representations directly from image datasets. However, deploying standard baseline models directly onto low-cost embedded elevator sensors remains difficult, as native networks struggle with passenger clutter, raincoats, mirror reflections, and the tight memory limits of edge microprocessors.

To overcome these deployment hurdles, the newly designed VGWS-YOLO model builds upon the compact You Only Look Once version 11 nano architecture, systematically modifying its feature extraction, bounding box regression, and post-processing stages. The researchers implemented three coordinated architectural enhancements: a VoV-Ghost Cross Stage Partial feature module, a Wise-IoU version 3 loss formulation, and a customized Soft Non-Maximum Suppression algorithm. This unified lightweight framework was tailored to handle the specific geometric and environmental characteristics of elevator cabins, maintaining real-time processing capabilities while minimizing computational footprint and parameter counts for embedded vision sensor nodes.

To train and evaluate the system under realistic operational conditions, the researchers constructed a custom surveillance dataset comprising 4,864 annotated images extracted from real-time monitoring cameras across residential districts and commercial office buildings in major Chinese cities, including Beijing, Shanghai, Shenzhen in Guangdong, Fuzhou, and Xiamen. The source cameras captured video at resolutions between 720P and 1080P and frame rates between 15 and 30 frames per second, observing cabins from mainstream manufacturers such as Otis, KONE, Schindler, Shanghai Mitsubishi, Hitachi, TKE, Toshiba, Fujitec, Yungtay, and Kangli. The dataset spans challenging real-world scenes, including varying cabin lighting conditions like direct glare and dim backlighting, complex reflections off polished metal walls, and varied e-bike appearances equipped with windproof quilts, raincoats, or storage trunks.

To eliminate temporal data leakage between nearly identical adjacent surveillance frames, the investigators partitioned the raw video clips using a clip-level disjoint splitting strategy rather than standard random frame-level sampling. Video clips were shuffled and partitioned at an eight-to-one-to-one ratio, yielding 3,831 training images containing 3,028 bicycle instances, 508 validation images containing 398 instances, and 525 test images containing 411 instances. All frames originating from the same recorded video sequence remained confined strictly to a single subset. A three-stage data augmentation pipeline—incorporating geometric transformations, photometric jittering and noise injection, and mixed occlusion simulation—was further applied to expose the network to diverse target orientations and pedestrian-vehicle interactions during training.

The structural enhancement within the network backbone centers on the VoV-Ghost Cross Stage Partial module, strategically inserted at Layer 19 of the neck network in place of the native C3k2 unit. Within the YOLOv11 architecture, Layer 19 produces a P4-scale feature map downsampled sixteen times, providing an effective receptive field suited for identifying medium-sized objects occupying central cabin space. The VoV-GSCSP module incorporates Cross Stage Partial channel splitting to reduce gradient redundancy, stacks lightweight GSBottleneck units using depth-wise and grouped convolutions for efficient multi-scale extraction, and aggregates intermediate feature maps in a single pass using One-Shot Aggregation. A final channel shuffling step promotes feature reuse while GSConv operators actively suppress specular reflections from elevator walls, preserving discriminative wheel and frame contours without overwhelming embedded memory bandwidth.

To refine spatial localization when vehicles are partially obscured by passengers or clothing, the team integrated the Wise-IoU version 3 loss function into the bounding box regression branch, replacing the baseline complete intersection-over-union metric. Unlike earlier static loss variants, Wise-IoU version 3 implements a dynamic non-monotonic focusing mechanism governed by an outlier degree parameter, calculated using the ratio of the instantaneous intersection-over-union loss to its exponential moving average across training iterations. By modulating gradient gains through dynamic hyperparameters set in this study to values of 1.9 and 3, the loss attenuates harmful gradient updates generated by heavily occluded outlier samples while concentrating optimization on medium-quality bounding boxes, avoiding the over-optimization of easily localized anchors.

For the final inference stage, the traditional greedy Non-Maximum Suppression algorithm was replaced with a customized Gaussian Soft-NMS method to prevent missed detections in cramped cabins where passengers and vehicles stand in close proximity. Standard hard-threshold suppression tends to inadvertently discard true positive detections when multiple adjacent objects share high spatial overlap, or retain duplicate bounding boxes when thresholds are relaxed. The Soft-NMS approach applies a continuous Gaussian decay function to smoothly reduce the confidence scores of overlapping predicted boxes based on their intersection-over-union ratios rather than instantly deleting them. Grid-search evaluations on the validation set determined that a decay parameter value of 0.5 optimized detection performance, after which only boxes maintaining adjusted confidence scores above 0.5 are preserved.

Controlled experimental benchmarking demonstrated that VGWS-YOLO achieved superior efficiency and detection accuracy compared against lightweight alternatives, including YOLOv5n, YOLOv6n, YOLOv8n, YOLOv10n, and native YOLOv11n. Benchmarked across three independent runs on a high-performance computing node equipped with an NVIDIA GeForce RTX 4090 GPU, VGWS-YOLO demonstrated a compact complexity profile of 2.428 million parameters and 6.098 GFLOPs. Forward inference throughput reached 157.38 frames per second under single-batch, FP32 precision settings. The model achieved a precision of 0.885, a recall of 0.953, a mean average precision at IoU 0.5 of 0.939, and an extended mAP50-95 metric of 0.679, outperforming the baseline YOLOv11n by 5.2 percent in recall and 3.3 percent in mAP50.

Ablation experiments confirmed the complementary nature of each architectural component within the full model. Independent integration of the VoV-GSCSP unit raised baseline precision from 0.869 to 0.889 and improved mAP50 to 0.924, though its stringent feature filtering slightly reduced recall to 0.883. Conversely, introducing Wise-IoU version 3 independently elevated precision to 0.891 and mAP50-95 to 0.666 by improving bounding box alignment. Implementing Soft-NMS on its own raised recall from 0.900 to 0.910 by recovering overlapping vehicle targets. Combining all three components achieved peak overall performance, where enhanced feature discriminability and regression stability countered individual trade-offs, driving recall up to 0.953 with narrow bootstrap confidence intervals confirming statistical reliability.

Field validation conducted across operating residential, commercial, and office building elevators confirmed practical operational feasibility under diverse real-world stresses, including motion blur, intense glare, and low cabin illumination. The deployed model maintained an average detection accuracy of 95.1 percent with an end-to-end processing latency of under 15 milliseconds, sustaining approximately 157.29 frames per second. The overall false-positive rate across non-target objects, such as luggage, baby strollers, and backpacks, was recorded at 2.1 percent, while visually similar pedal bicycles produced a false-positive rate of 2.4 percent. Because undetected electric bicycles present critical fire hazards in safety management, the authors emphasized that prioritizing high recall over minor false alerts from non-motorized bicycles remains the optimal trade-off for community monitoring.

The researchers noted operational limitations, observing that recognition accuracy degrades when objects face extreme obscuration exceeding 50 percent or lighting levels drop below 50 lux. Additionally, the dataset leaves room for broader sampling of rare vehicle modifications and unusual parking orientations. The authors noted that physical on-device inference speeds, power consumption, and thermal stability on embedded edge processors, such as ARM-based neural processing units and NVIDIA Jetson hardware, require further direct hardware measurement. Future investigations plan to explore context-aware feature fusion, adaptive illumination compensation, and expanded detection classes to monitor additional elevator hazards, including flammable liquids and open lighters.

Subject of Research: Lightweight computer vision model for electric bicycle detection in residential elevators

Article Title: An efficient visual sensing method for electric bicycle detection in complex elevator environments based on VGWS YOLO

Article References: Liu, Q. (2026). An efficient visual sensing method for electric bicycle detection in complex elevator environments based on VGWS YOLO. Discover Computing, 29(1), Article 861. https://doi.org/10.1007/s10791-026-10840-4

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10840-4

Keywords: Object Detection, Computer Vision, YOLO Architecture, Electric Bicycles, Elevator Safety, Edge Computing, Deep Learning, Thermal Runaway, efficient, visual, sensing, method