Preindication mining for predicting pedestrian action change

Abstract

The action prediction of pedestrians significantly contributes to an intelligent braking system in cars; knowing that the pedestrians will run in several seconds such as for crossing streets, the cars can start braking in advance, to effectively reduce the risk for crash accidents. In this paper, we propose a method to predict how the pedestrian act (run or walk) in the future based on preindication in video frames detected by only appearance-based image features. We empirically mine the distinctive frames that precede the target action, ‘running’ in this case, and are effective for predicting it in the framework of feature selection. By using the most effective frames, we can build the action prediction method by exploiting the image features extracted at those frames. As to the image feature extraction methods, we evaluate two types of features in our method, one is GLAC (Gradient Local AutoCorreration) and the other is HOG (Histogram of Oriented Gradient). In the experiments, the effective frames are successfully found around 0.37 s before running action by using GLAC feature; this is not the case of HOG. We also show that the results are closely related to human motion phases from walking to running via biomechanical analysis.