Multiple Object Motion Prediction Using Deep Convolutional Neural Networks

Abstract

This paper presents multiple object motion prediction using Deep Convolutional Neural Network (DCNN). Specifically our approach generates a potential map from the location data of multiple objects, and the DCNN learns to predict the future potential from the previous time frame. Advantages of this model are enabling the behavior prediction without using tracking by acquiring each object’s successive state. In addition, it was confirmed in our experiment that our model enables the multiple object detection at about 6 msec. per frame using GeForce GTX TITAN X of NVIDIA’s GPU.