3D SEMANTIC SEGMENTATION USING CAMERA-LIDAR SENSOR FUSION FOR AUTONOMOUS DRIVING

Abstract

This chapter we discussed the results from our experiments on the Kitti dataset qualitatively and quantitatively. The experiments and evaluation methods for different architectures for the middle fusion and early fusion compare them with the ground truth in the mean-IOU metric. We showed that we had enhanced 10% of the accuracy by raising the overall detection mIOU of the early fusion and middle fusion 36.7% and 37.4%, respectively. On the other hand, the no-fusion architecture mIOU of 33.7% in SqueezeSeg architecture and scored 37.8% and 37.6% respectively instead of 34.8% in PointSeg architecture, showing the advantage of fusing the data in early fusion and middle fusion instead of using single sensor modality. In the next chapter, we will discuss the conclusion of our work, experiments, and analysis that can be done in the future.