Real-Time Helmet Detection and Automatic Number Plate Recognition for Motorcycle Traffic Surveillance Using YOLOv3 and CNN

Abstract

Motorcycle fatalities are a pressing public health issue worldwide. A lack of helmet usage contributes to a disproportionately high percentage of riders killed in traffic collisions. Conventional enforcement is expensive, spatially limited, and by nature, reactive. In this paper, we present a real-time, automated pipeline for helmet compliance detection and number plate recognition from motorcycle traffic videos. A custom trained YOLOv3 model detects and localizes riders and motorcycles; a Convolutional Neural Network (CNN) binary classifier recognizes non-compliant riders; and a violation-conditioned Optical Character Recognition (OCR) module performs license plate text extraction only on non-compliant riders , thereby saving 22% of computation. No work to date has combined detection, classification, conditional ANPR and web deployment into a single integrated pipeline; this is the main contribution of this paper. We provide the system via an interactive Streamlit web application that does not require any client side software. Evaluation on a 420 frame test set shows a detection mAP of 88.3%, a helmet classification accuracy of 93.7% and an 89.4% character level plate recognition accuracy.

Introduction

The text presents a deep learning–based real-time system for automated helmet violation detection and enforcement in traffic surveillance. It addresses the high global rate of road fatalities, especially among motorcyclists, and emphasizes that helmet use significantly reduces head injuries. However, manual enforcement methods are limited due to traffic volume, labor constraints, and delays in CCTV analysis, motivating the need for an automated solution.

The proposed system integrates three components into a single pipeline: (1) YOLOv3 for detecting motorcycles and riders, (2) a CNN classifier for determining helmet usage, and (3) a conditional ANPR (Automatic Number Plate Recognition) module using Tesseract OCR, which is triggered only when a rider is detected without a helmet. This conditional design improves efficiency by reducing unnecessary OCR processing and lowering latency. The system is deployed using a Streamlit web application, enabling easy real-time use without specialized hardware.

YOLOv3 performs object detection using multi-scale predictions, while the CNN classifies helmet compliance using cropped rider images. If a violation is detected, the system isolates the license plate region through image processing techniques and extracts text using OCR.

The study also reviews related work, highlighting earlier approaches such as traditional image processing, HOG-SVM models, and deep learning detectors like Faster R-CNN and YOLO variants. While prior systems achieved good accuracy in detection, most lacked integrated license plate recognition or deployable end-to-end systems. The proposed work improves on this by combining detection, classification, and OCR into a unified architecture.

Conclusion

This paper presents a real-time, violation-conditioned deep learning pipeline for motorcycle traffic monitoring. The system integrates custom YOLOv3 detection, CNN helmet classification, and OCR-based plate recognition in a single deployable Streamlit application. Experimental evaluation yields 88.3% mAP, 93.7% classification accuracy, and 89.4% plate character accuracy, with a 22% per-frame latency reduction attributable to conditional OCR gating. The proposed system is the first unified detect–classify–ANPR–deploy pipeline for this application domain. Future work targets four directions: (1) migrating the detection backbone to YOLOv8 for improved speed-accuracy trade-off; (2) replacing Tesseract with a CRNN-CTC plate recognizer for robustness under blur and distortion; (3) edge deployment on NVIDIA Jetson modules for infrastructure-free field installation; and (4) integration with smart city alert systems to issue real-time violation notifications and automated e-citations to traffic control centers.

Copyright

Copyright © 2026 Kundan Kumar Patel, Abhishek Kumar, Vikas Gupta. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.