Using 3D Object Detection DNN in an Autonomous Tram to Predict the Behaviour of Vehicles in the Road Scene

In this paper we propose a solution to the problem of predicting the behavior of vehicles on the road scene based on image analysis. A 3D detection deep neural network (DNN) is used to detect objects and estimate their positions. The kinematic bicycle model is considered as a behavior model, with each vehicle being processed separately, i.e. without taking into account the influence of other road agents. This paper provides detailed information about the detection network in use with our training approach, as well as the obtained metrics reflecting the accuracy of output navigation parameters. In conclusion, the article presents the accuracy characteristics that determine the time of the "reliability" of the model forecast, obtained on an open Agroverse dataset.