Convolutional Neural Network Information Fusion based on Dempster-Shafer Theory for Urban Scene Understanding

Dempster-Shafer theory provides a sensor fusion framework that autonomously accounts for obstacle occlusion in dynamic, urban environments. However, to discern static and moving obstacles, the Dempster-Shafer approach requires manual tuning of parameters dependent on the situation and sensor types. The proposed methodology utilizes a deep fully convolutional neural network to improve the robust performance of the information fusion algorithm in distinguishing static and moving obstacles from navigable space. The image-like spatial structure of probabilistic occupancy allows a semantic segmentation framework to discern classes for individual grid cells. A subset of the KITTI LIDAR tracking dataset in combination with semantic map data was used for the information fusion task. The probabilistic occupancy grid output of the Dempster-Shafer information fusion algorithm was provided as input to the neural network. The network then learned an offset from the original DST result to improve semantic labeling performance. The proposed framework outperformed the baseline approach in the mean intersection over union metric reaching 0.546 and 0.531 in the validation and test sets respectively. However, little improvement was achieved in discerning moving and static cells due to the limited dataset size. To improve model performance in future work, the dataset will be expanded to facilitate more effective learning, and temporal data will be fed through individual convolutional networks prior to being merged in channels as input to the main network.