Abstract
With the increase of roads and vehicles, the number of traffic accidents has gradually increased, and one of the important reasons is the driver’s unsafe driving behavior. Some existing methods rely on the method of joint point estimation to identify driving behavior. This kind of method can filter out unrelated components such as background, but the joint point estimation can only give the position information of human behavior, which cannot fully describe the human body, so we try to solve this problem. For one thing, we propose a driver behavior recognition method based on human parsing which can quickly and accurately recognize the driver’s unsafe behavior. For another thing, we build a trucker behavior dataset (TB dataset) with 700 videos and improve a human body segmentation model with higher precision. Finally, we achieve 93.39% and 91.32% accurate for recognizing driving with one-hand and no-hand. At the same time, we get 98.75% accuracy for recognizing driving with turned head by using ResNet50 model.
Highlights
MOTIVATION AND OBJECTIVE As it is widely acknowledged that road traffic accidents, in which human being plays an essential role, are serious issues that faced by the whole world
Statistics from the Traffic Management Bureau of the Ministry of Public Security of the People’s Republic of China show that in 2016, there were a total of 50,400 road traffic accidents involving trucks, resulting in 25,000 deaths and 46,800 injuries
Due to the great success of convolutional neural network [5] for computer vision area, many recent works use depth models to train networks to identify behavior in video and the performance of these new proposed methods is significantly better than hand-crafted methods [6], [7]
Summary
Deep learning is a data-driven approach, so in order to solve the problem of unsafe driving, we need to build the dataset to train and test our network. We propose a human parsing network to obtain the segmentation results of all images. We judge and analyze behavior of large truckers based on the results of the above human parsing.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
