An Algorithm for Detecting Traffic Control Hand Signals

Abstract

The development of autonomous driving technology is accelerating. In particular, there are many studies on situations that appear while driving. The situation in which the police use hand signals is also an example. In foreign countries, there have been several attempts to recognize police hand signals by using depth images, skeleton extraction, and acceleration sensors. However, these methods are computationally intensive and have limitations that only work in limited circumstances. These studies have not even achieved real-time of 30 fps. Therefore, in this paper, an algorithm for recognizing and interpreting traffic control hand signals using an object detection algorithm was developed without using sensors other than the camera. The hand signal detection system consists of an RGB detector that detects the classes in the arm direction of the police officer , and an RNN-based classifier that interprets the arm direction class sequence as go straight, left turn, right turn, stop, and no signal. In this study, yolov4 was used as an RGB detector, and RNN algorithms such as Vanilla RNN, LSTM, Bi-LSTM, and GRU were compared as classifiers. The accuracy was compared by varying the number of hidden layers. As a result, the GRU of 4 hidden layers showed high accuracy and fast speed, and was appropriate as a classifier. In classifying the hand signal, it was more accurate to set the condition as a threshold value than to use the maximum value. A method of comparing the modes of subsequences was also proposed to distinguish valid and invalid signals. As a result, a speed of 30 fps was achieved in the FHD video.