An Algorithm for Detecting Traffic Control Wand Signals

Abstract

In the 3rd to 5th stages of autonomous driving, the vehicle must be aware of the external environment and control itself.[1] In the case of a formal situation on the road, it is not difficult to perceive the situation by itself. However, it is difficult to judge in informal situations such as traffic construction, congestion, accidents etc., in these case, the vehicle must recognize and classify the signal of the foreman. In particular, since the foreman uses a wand for visually effective signal indication, it is necessary to determine the trajectory of the wand to control. In this study, computer vision and deep learning-based algorithms are applied to detect foremen and wands, and the information of the foremen and wands is input to the Recurrent Neural Network (RNN) for classifying the go/no signal/front stop/side stop/turn left/turn right. As a result of experiment by modifying the number of layer, the average accuracy was 93.55% and the average 28.78 fps was obtained in the case of 2 hidden layers, and experiment by modifying time step (16/24/32) shows high accuracy for the case of 24 time step (93.36%). Especially, there is a significant difference in accuracy at the case of ‘turn left/right’. Finally, the suggest selected input features was 12.53% more accurate than simple coordinates.