Temporal-Aware Convolutional Network for First-Person View Pedestrian Trajectory Prediction

Pedestrian trajectory prediction is one of the key technologies to realize safe and reliable autonomous driving systems. Compared with the traditional bird's-eye view dataset, pedestrian trajectory prediction from the first view can provide direct visual information about pedestrians, which has higher application value. In this study, a multi-modal pedestrian trajectory prediction model with time-aware convolutional network (TACN) is proposed. By transforming the raw trajectory data into a multivariate input format, the model can more precisely capture both the pedestrians' location information and their dynamic changes. At the same time, the designed time-aware convolutional network can extract multi-time scale feature, and effectively deal with the immediancy and burstiness of pedestrian behavior. To deal with the uncertainty of future trajectories, we employ a conditional variational autoencoder (CVAE) to implement multimodal trajectory prediction and generate multiple possible future trajectories. Experimental results demonstrate the superiority of our approach.