Investigation of Action Recognition for Improving Pedestrian Intent Prediction

Pose estimation has been a critical aspect for the recent improvements made in the field of pedestrian intent prediction. Current pose estimators are capable of providing highly accurate posture and head orientation information. In our previous work, we utilised posture information for predicting the crossing behaviour of pedestrians in urban environments. We referred to this as the multi-scale pedestrian intent prediction (MS-PIP) architecture. This technique yielded state-of-the-art results of 94% accuracy. It has been suggested from some previous works that head orientation information provides insight into the pedestrian’s behaviours and intentions. Therefore, in this study, we investigate the benefits of implementing head orientation on top of the existing MS-PIP architecture. We found that the addition of head orientation information in fact decreases accuracy when compared to our previous works, in some cases by over 50%. Data augmentation and data generalisation techniques were also applied which slightly improved the accuracy. However, the accuracy was still lower than the original MS-PIP results.