Jul 2026· 2026 5th International Conference on Distributed Computing and Electrical Circuits and Electronics (ICDCECE)· pp. 1-8· 0 citations· 20 references
Abstract
Motorcycle accidents caused by helmet non-compliance remain a major public safety concern, particularly in densely populated urban environments where manual traffic enforcement is insufficient. This paper presents a distributed Edge AI-based rider safety framework capable of real-time helmet detection and automatic speed restriction. The proposed architecture separates visual inference and motor actuation into two independent computational nodes using a Raspberry Pi 5 and RP2040 microcontroller respectively, thereby minimizing control latency and improving deterministic response reliability. An optimized YOLO11s object detection model was trained using a custom dataset containing helmet, no-helmet, and adversarial edge-case samples such as caps and partially occluded faces. The trained model was quantized using the NCNN framework for efficient ARM-based edge deployment. Experimental evaluation demonstrated a best mAP@50 of 0.9034, precision of 0.9248, and average inference latency of approximately 45 ms, corresponding to nearly 22 FPS on Raspberry Pi 5 hardware. To reduce false-positive activations, a five-frame sequential validation mechanism and hysteresis-based PWM throttle regulation strategy were implemented. The system successfully performed real-time throttle limitationduring no-helmet detection scenarios while maintaining stable unrestricted operation for compliant riders. The proposedframework demonstrates the feasibility of combining Edge AI perception with hardware-level safety intervention for intelligent two-wheeler rider protection systems.
This paper proposes a real-time monitoring and warning system for dangerous electric bicycle driving behaviors based on the YOLOv11 model and the K230 edge computing platform. The system is designed for urban roads and campus traffic environments, where behaviors such as riding without helmets, illegal passenger carrying, and speeding frequently occur. Compared with traditional cloud-based monitoring methods, the proposed system can perform real-time detection on edge devices with lower deployment cost and lower latency. The system uses the K230 edge AI chip based on the dual-core RISC-V architecture [6] as the main hardware platform and deploys the lightweight YOLOv11 model for traffic violation detection. To improve detection performance for small targets such as helmets and partially occluded riders, the C3k2 and C2PSA modules were introduced into the model structure. In addition, multiple data augmentation methods were used during training to improve robustness under different lighting and weather conditions. To improve the accuracy of monocular speed estimation, this paper introduces a speed measurement method based on camera calibration and inverse perspective mapping (IPM). Compared with traditional pixel-displacement methods, the proposed method reduces the influence of perspective distortion and improves measurement accuracy without additional radar sensors. Furthermore, a multi-frame decision strategy based on temporal consistency is used to reduce false alarms caused by temporary occlusion or unstable detection results. Experimental results show that the INT8-quantized model achieved an inference speed of 22 FPS on the K230 platform, while maintaining 91.9% mAP@0.5. The average relative error of speed estimation was controlled within 5.3%. The proposed system combines object detection, speed estimation, and remote warning functions into a complete edge-side monitoring framework, providing a practical solution for intelligent traffic supervision.
Mengting Zhang, Chengjian Dong· International Conference on...· 0 citations
This paper presents a machine learning pipeline for real-time crash and fall detection using only smartphone accelerometer and gyroscope data, without dedicated hardware, and proposes a tiered escalation architecture to absorb the false-positive cost while preserving high sensitivity to genuine falls.
Onkar Atul Allewar· International Journal for Re...· 0 citations
The proposed system is an ADAS-integrated safety system to reduce road accidents caused by impaired driving conditions such as alcohol influence and driver fatigue. The system uses a dual controller architecture, where an ESP32 micro-controller and a Raspberry Pi are combined to achieve efficient and real-time operation. The ESP32 mainly handles the safety critical functions. These include alcohol detection, vehicle speed monitoring, engine lockout, and adaptive speed control. On the other hand, the Raspberry Pi is used for vision-based drowsiness detection. It performs facial landmark analysis to identify signs of fatigue from drivers. To further improve the system capability, GPS based tracking is also included with the cloud supported telematics. This enables real-time fleet monitoring and alert generation. During critical conditions, vehicle location and alerts information is sent to the authorized team. At the same time, the driver receives safety warnings and next safe rest stop recommendations when fatigue is detected. The proposed system combines driver monitoring, vehicle control, and telematics into one embedded solution. As a result, it can be applied effectively in intelligent transportation and fleet safety systems.
A. S, Vijendra Babu D· 2026 International Conferenc...· 0 citations
This research method outperforms existing mainstream models in terms of accuracy, efficiency, and interference tolerance, providing reliable technical support for real-time driving behavior monitoring.
Guozhu Sui, Meixia Song, Hai-Yun Sun et al.· Information Technology and C...· 0 citations
Collisions between motorcycles and opening car doors are common causes of motorcycle-related accidents. To avoid these collisions, this study proposed a new crash alert system embedded in motorcycles to warn motorcyclists of opening car doors ahead, using artificial intelligence and a matching algorithm. The proposed method involved two stages: (1) training a deep learning model to detect car door states and (2) applying a matching process to eliminate incorrect opening door predictions to improve object detection model performance. Experimental results demonstrate that the YOLOv12s object detection model achieved strong performance in predicting car door states, with precision, recall, and mean average precision (mAP) values of 85.5%, 84.0%, and 88.8%, respectively. Validation on the test dataset further confirmed the model’s effectiveness, yielding precision and recall values of 86.8% and 83.6%, respectively. Moreover, integrating the matching algorithm improved precision to 92.0%. These findings highlight the high feasibility of the proposed method for real-world applications.
Y. Shiao, Tan-Linh Huynh· E3S Web of Conferences· 0 citations
To meet the demand for real-time detection of electric bicycle helmet wearing in intelligent transportation scenarios, this paper constructs a helmet recognition and detection model based on YOLOv8. A self-built traffic scene dataset containing 5,448 images is established, covering complex road conditions including multiple weather conditions, multi-view angles, far and near shooting distances, day and night environments. The dataset is split into a training set and a validation set at an 8:2 ratio. All model training experiments are carried out on hardware equipped with an Intel i9-13900K CPU and an NVIDIA RTX 4090 GPU, with a batch size of 16, an initial learning rate of 0.0001, and a maximum training epoch of 100. Precision, Recall, F1-Score, mAP@0.5 and mAP@0.5:0.95 are adopted as core evaluation metrics. Experimental results show that the overall mAP@0.5 of the model reaches 0.911; the maximum precision hits 0.972 at a confidence threshold of 0.972, and the global Recall is 0.95. The optimal F1-Score of 0.87 is achieved at a confidence threshold of 0.443. All metrics converge and stabilize after 50 training epochs. When deployed on an NVIDIA GTX 1060 graphics card, the model achieves a single-frame inference latency of 33 ms and a frame rate of 30 FPS, which satisfies the real-time requirement of 25 FPS for traffic monitoring systems. The proposed model exhibits strong robustness under variable illumination, viewpoints and weather conditions, providing a high-precision, low-latency visual detection solution for intelligent supervision of electric bicycle helmet compliance.
Yi Yao, Hui Zhang, Jianjun Xiang et al.· IEEE International Conferenc...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
