Pedestrian Intent Prediction by Dense Graph Transformer
Oscal Tzyh-Chiang Chen, Han‐Wen Liu, Yu-Lung Chang, Yu-Wei Jhao · 2022 IEEE International Conference on Consumer Electronics - Taiwan · 2022
In this work, we develop a Deep Neural Network (DNN) to predict pedestrian intent in a road where two input streams of subject images, and their skeletal joint points, bounding box points and bottom middle point of a scene are provided. The proposed DNN includes 3D convolutional neural network, self-attention layer, reduction layer for processing subject images, and dense graph convolution or dense graph transformer for skeleton-position data. The proposed DNN was verified on Joint Attention for Autonomous Driving (JAAD) database and achieved an accuracy of 82.34%. As compared to the conventional work [4], the proposed DNN reveals superior performance on the task of pedestrian intent prediction.