Distracted Driving Detection in Low-Quality Images Using Deep Learning

Research aim:Considering increased rates of car accidents in our country, there is a need to supervise, damage control and present an innovative system or algorithm in the field of computers for detecting different types of driving violations such as talking on the phone, eating and drinking, etc. Our goal is to distinguish and classify driving violations. Regarding the reasons for doing this research is the disability of the classic methods for distinction of the driving violations because of the excessive amount of the data. Nowadays we can use various methods and ways in the field of artificial intelligence and machine vision for the distinction and classification of the driving violations. We have utilized deep learning in this research for its accuracy in distinction of objects in the images and also its classification of the data. Research method:In this research, first we study the classic methods in distinction of the driving violations, then we study machine learning, deep learning and its subdivisions, methods and tools in deep learning and the convolutional neural networks. In deep learning one of the algorithms for distinction and classification of the images is the convolutional neural network with various architectures of the networks. Findings:We have introduced a new architecture which in addition to the improved VGGNet, has less parameters compared to VGGNet architecture , has architecture of distinguishing the images without noise, and also has acceptable function in distinguishing images with intense noise in 10 classes (10 types of driving forms). Conclusion:The total number of our suggested architecture parameter is nearly 72% less than original architecture of VGGNet-19 which leads to reduction in the compute and processors at the network. The results of our suggested architecture experiments were 99/22% of accuracy in the images without noise and 99/12% in detecting and classifying the images with noise.