Abstract
In order to predict pedestrians' movements and help the pedestrian vehicle interaction, it is important to make a computer vision based classification for whether the pedestrian is looking at the vehicle. Conventional methods based on head orientation estimation only estimates the general orientation of the pedestrian head. Besides, the results are not good due to the low resolution of the pedestrian head image. Meanwhile, a large number of new features are extracted from the head images, which would increase the computation complexity. In this thesis, we propose a novel method that can predict the pedestrian's attention. The proposed method classifies whether the pedestrian is looking at the vehicle, using human pose estimation. The algorithm of proposed method consists of the following modules: (1) A subset of JAAD data set is selected as the dataset for this research. (2) As an existing estimation tool, Alphapose is used to obtain key points on the pedestrian's body. (3) In two successive frames, the pedestrian is tracked by Intersection over Union (IOU). (4) Pedestrian features based on the position relationship and confidence of key points in the successive frames are extracted from the result of Alphapose. (5) The model trained by random forest algorithm is used to classify the pedestrian's attention. Experimental results show that the proposed method can achieve a high accuracy for classifying the pedestrian attention.
