Conflict Management in a Distance to Prototype-Based Evidential Deep Learning

In the domain of autonomous vehicles, perception tasks are very complex, and deep learning can be coupled with evidence theory for uncertainty management of perception models. If the pieces of evidence involved in the merging process of the deep learning-based model are discordant, the results can be degraded. Therefore, verifying the conflicting level of sources and alleviating it when possible, gives the capability to increase efficiency of following steps of the workflow: fusion rules and decision making. This paper highlights scenarios where high conflict values occur within an evidential neural network architecture. The cause of conflict is analyzed and a conflict management method is proposed, allowing the appropriate use of fusion rules and decision-making based on belief functions. Thus, particularly in, the distance to prototypes approach, a parameter rectification is proposed. The experimental results are obtained using a lidar-camera cross-fusion architecture with evidential formulation based on Dempster-Shafer’s theory. The model is investigated on a road detection task and it uses the KITTI dataset.