レーザセンサによる行動モニタリングデータを用いた機械学習による歩行者軌跡分類法

抄録

 The trajectory analysis of pedestrians using the tracking system of laser scanners is an effective approach for understanding the usage of facility spaces. Carefully looking at the pedestrian trajectories observed in an outpatient reception area of a hospital, we can find some specific patterns of patient behavior. For instance, some patients come in the area from the main entrance, line up at the return visit reception machines, and leave to the consulting room area slowly. Also, some coming from the main entrance, pass through the return visit reception machines, and leave the consulting room area quickly. These examples indicate that behavior patterns and attribute information of facility users can be estimated from their trajectory features such as directionality, staying place and walking speed. However, it requires a heavy load to manually extract the features from pedestrian trajectories and classify them into some adequate patterns. Hence, it is highly desirable to achieve this task automatically.
 In this paper, we proposed a method of pedestrian trajectory classification using the Restricted Boltzmann machine, by which we can automatically find the inherent features of pedestrian trajectories. This method was applied to an outpatient waiting area of a hospital. Comparing manual and automatic classification, we demonstrated the usefulness and sufficient performance of our proposed method in extracting the feature of directionality, staying place and walking speed. The details are as follows:

 (1) Modeling of pedestrian trajectory
 The trajectory data were divided into three-layers composed of 1 m square grids, which were consisting of “Front layer”, “Back layer” expressing the directionality, and “Staying layer” expressing staying places. The restricted Boltzmann machine had input units and binary hidden neurons, by which the feature of the trajectory data were generated after sufficient learning. In setting the number of hidden neurons, the 100 × 8 model, which had 100 neurons in the first hidden layer and 8 neurons in the second hidden layer, was applied by comparing the information entropy of the hidden layer.

 (2) Model validation
 In the 100 × 8 model, the degree of coincidence between the results by manual classification and automatic classification was examined. The entropy ratio, which is an index for checking the degree of agreement, was 0.6% in the entropy ratio by manual classification, and 10.6% by automatic classification. The results showed that the manual classification and the automatic classification was in good agreement. In addition, the trajectory distribution diagrams were configured for each machine classification, and the feature pattern diagrams were made by 2nd hidden neurons, which automatically found inherent features. These diagrams demonstrated the effectiveness of our proposed method.