ResNet-Based Model for Autonomous Vehicles Trajectory Prediction

Abstract

Autonomous vehicles (AVs) are expected to dramatically redefine the future of traffic. However, there are still plenty of challenges need to be figured out before L5 self-driving era coming. One of them is to precisely predict the moving trajectory of traffic agents which near the AV, such as cars, pedestrians, and motorcycles. In this paper, we use ResNet to forecast AVs' trajectories, which is able to capture the features of different dimensions to achieve better predictions. By feeding the raw input picture, the model output s three trajectories and their confidence levels respectively, which means each trajectory has its own confidence level. Experimental results show that our method performs better than other deep learning methods. The loss function value of ResNet-34 model is lower than that of VGG-16 model and VGG-19 model.