Exploring image recognition: applying convoluted neural networks and learning to recognize safe cyclists

Neural networks;Convolutional neural networks;Road safety;Image recognition

Narayanan, Ramakrishnan ; Sowers ; Richard B.

关键词: Neural networks;    Convolutional neural networks;    Road safety;    Image recognition;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/99141/NARAYANAN-THESIS-2017.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Today, there is a need to focus on the mobility revolution that is currently taking place. With the advent of more intelligent data gathering, there is also a growing need for using existing technology and infrastructure to achieve this goal, without incorporating expensive, complicated systems. As single-occupancy give way to shared mobility solutions, combined with regular mass transit and pedestrian-aware street infrastructure (traffic lights, crosswalks etc.), there is a large "networked mobility system'' that has the potential to be tapped. Moreover, autonomous cars will be here soon, to add to the mix. With statistics showing an increase in bicyclist related crashes over the last decadeand an increase in bicycle-borne road users, there is a necessity for cities and autonomous vehicles to build bicycle safety into their adaptation to the "driverless future". This paper is an exploration into the use of a Convolutional Neural Network (CNN) based Machine Learning (ML) algorithm to identify bicycle-borne road users, who wear helmets. We use a pre-made CNN framework-YOLO (You Only Look Once), and built around it further. After a brief proof-of-concept test on a publicly available dataset (including extraction, parsing and detection), the algorithm was modified. Some important features were added, such as identifying license plates, faces and encrypting them. Further, there is also a detailed account of using the ML capabilities that the framework is built with, and training it to identify bicycle-borne road users wearing a helmet.