Research on Behavior Recognition of Disyielding Pedestrians Based on LSTM

Abstract

Serious traffic accidents caused by vehicles' disobedience to pedestrians occur frequently, so it is very important to supervise and regulate the behavior of disobedient pedestrians and explore the identification methods of disobedient behaviors. Compared with the existing research, this paper proposes a recognition algorithm based on Long Short Memory (LSTM) network for vehicles' indecisive pedestrian behavior. This algorithm is more practical and adaptable. First, it predicts the future continuous kinematics of the vehicle through the past transverse and longitudinal characteristics of the vehicle, and obtains the driver's manipulation intention. Then, the manipulation intention is combined with the past pedestrian trajectory and pedestrian head direction information, and the joint modeling is carried out to obtain the probability of the occurrence of vehicles' behavior that does not yield to pedestrians. Using real driving data to test, the test results show good recognition accuracy, and the goodness of fit of the model to the sample reaches 0.93.