Multi-Agent Cooperative Camera-Based Semantic Grid Generation

The idea of cooperative perception for navigation assistance was intro-
duced about a decade ago with the aim to increase safety on dangerous
areas like intersections. In this context, roadside infrastructure appeared
very recently to provide a new point of view of the scene. In this
paper, we propose to combine the Vehicle-To-Vehicle (V2V) and Vehicle-
To-Infrastructure (V2I) approaches in order to take advantage of the
elevated points of view offered by the infrastructure and the in-scene
points of view offered by the vehicles to build a semantic grid map of
the moving elements in the scene. To create this map, we chose to use
camera information and 2-Dimentional (2D) bounding boxes in order to
minimize the impact on the network and ignored possible depth informa-
tion as opposed to all state-of-the art methods. We propose a framework
based on two fusion methods: one based on the Bayesian theory and the
other on the Dempster-Shafer Theory (DST) to merge the information
and chose a label for each cell of the semantic grid in order to assess
the best fusion method. Finally, we evaluate our approach on a set of
datasets that we generated from the CARLA simulator varying the pro-
portion of Connected Vehicle (CV) and the traffic density. We also show
the superiority of the method based on the DST with a gain on the
mean intersection over union between the two methods of up to 23.35 %.