International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 09 Issue: 07 | July 2022
p-ISSN: 2395-0072
www.irjet.net
Vision-Based Motorcycle Crash Detection and Reporting Using Deep Learning Sanjit Gawade1, Nadine Dias2 1Student,
Information Technology Department, Goa College of Engineering, India Professor, Information Technology Department, Goa College of Engineering, India ---------------------------------------------------------------------***--------------------------------------------------------------------based on GPS and GSM technology, and ultrasonic and Abstract - In recent years, two-wheelers have gained 2Assistant
impact sensors inside cars to detect and report the occurrence of an accident. Researchers are working on vision-based accident detection, it was observed that those systems were trained using cars and heavy vehicle data. Such systems are based on finding acceleration anomaly, trajectory anomaly, and change in angle anomaly [1], this approach is effective to determine car accidents in normal traffic flow and good visibility conditions and requires postprocessing of data.
popularity amongst daily commuters, youngsters, and urban residents as two-wheelers are easy to operate in congested traffic conditions and are fuel-efficient. With the increase in the number of vehicles on road, there has been a significant increase in road accidents. When we compare safety features offered on vehicles, motorcycles are equipped with the least features compared to cars and other vehicles. This project endeavors to provide a means to explore a vision-only approach to detecting traffic anomalies. Considering the realtime operating aspect of the system, the YOLOv4 algorithm is considered for object detection and classification. The algorithm is trained using a custom dataset with 398 images of road anomalies involving a motorcycle. The Model performed exceptionally well and achieved mAP@50 of 74% and precision of 60%.
Our proposed system emphasizes a vision-only approach to detect anomalies on road. The main focus is to detect motorcycle accidents, as there are fewer safety features on motorcycles, and injuries attained may prevent a wounded person from contacting emergency medical services and receiving medical attention. This research endeavors to provide a means to explore a vision-only approach to detecting traffic anomalies. The main focus is to detect incidents involving motorcycles, as there are fewer safety features on motorcycles, and injuries attained may prevent a wounded person from contacting emergency medical services. Considering the real-time operating aspect of the system, 3 algorithms were finalized Faster-RCNN, Yolov4, and Yolov4-Tiny. After training the models, it was found that the YOLOv4 algorithm Outperforms Faster R-CNN in terms of speed and Yolov4-Tiny in terms of accuracy. The algorithms were trained using a custom dataset with 398 images of road anomalies involving a motorcycle. The Yolov4 Model performed exceptionally well and achieved mAP@50 of 74% and precision of 60%.
Key Words: Deep Learning, Computer Vision, Yolov4, Object detection, Real-time, CNN. 1. INTRODUCTION One of the most compelling types of artificial intelligence is computer vision. Computer vision is the field of computer science that focuses on understanding complex human vision systems and enabling a computer to identify and process information in images and video streams like humans do. The application of computer vision is expanding in fields like healthcare, surveillance, action and activity recognition, military, agriculture, and manufacturing. Computer vision algorithms analyze certain criteria in images and videos and then apply interpretation to predictive or decision-making tasks. Computer vision models are designed to translate visual data based on feature and contextual data learned during training.
2. Related work The evolution of object detection started in the early 2000s. They followed the low and mid-level vision and followed the method of recognition by components. This method enabled object detection as a measurement of similarity between the object components, shapes, and contours and the feature that were taken into consideration were shape context, edgelessness, and distance transformation.[doi2000]. object detection genre was not making any progress as the performance of hand-crafted features became saturated. However in 2012, with the advancement in convolutional neural networks and deep convolutional networks, they were successful at learning robust and high-level feature representations of an image. The deadlocks of object
Convolutional Neural Network is the foundation of modern computer vision algorithms. CV algorithms are based on CNN, which provides a dramatic performance improvement compared to traditional image processing algorithms. CNNs are neural networks with a multi-layered architecture that is used to gradually reduce data and calculation to the most relevant set. After the initial survey of automated accident detection systems, it was identified that existing systems are mainly
© 2022, IRJET
|
Impact Factor value: 7.529
|
ISO 9001:2008 Certified Journal
|
Page 2752