Trajectory Prediction Using Video Generation in Autonomous Driving

Trajectory prediction for the surrounding cars is a useful task in autonomous driving for obvious reasons. The traditional methods for predicting the future trajectories of surrounding cars involved complex motion models and patterns, complex maneuvers or physical models of the car trajectories. More recent works aim to predict the future car positions by using deep learning and neural networks. In this paper, video generation models were employed, which provide an estimation of the future frames related to the car positions based on an existing video and can obtain the position of the selected cars by employing an object detection algorithm along with additional information obtained by a segmentation module that uses a semantic segmentation network. The results were validated by employing the Root Mean Square Error (RMSE) metric in order to predict the locations of the surrounding cars and estimate their depth. Apparently, this approach has never been implemented in order to obtain the trajectory and the future position of the surrounding cars in autonomous driving.