Crowd Trajectory Prediction Based on Surveillance Data

Abstract

In recent years, with the increasing flow of people in public places, people's awareness of safety and personal space have continuously increasing. Therefore, a large number of surveillance cameras are installed in public places, generating a large amount of surveillance data. Trajectory prediction based on the surveillance data has become a hot research direction in the field of computer vision and crowd behavior analysis. It can be used in traffic control and management, road planning, prevention of safety accidents and many other critical domains. The purpose of this paper is to propose a more efficient trajectory prediction algorithm. This paper proposes to use the convLSTM (Convolutional Long Short-Term Memory) neural network to predict pedestrian trajectory, and obtains better results than its original model LSTM (Long Short-Term Memory) network in the acquisition of space-time relationships. Combined with the mixed density network (MDN), the network model uses a probability distribution instead of a definite location to predict the trajectory, which greatly improves the efficiency and accuracy of trajectory prediction.