Automotive Interior Sensing - Imaging Solutions

Abstract

This dissertation addresses the problem of object detection inside the vehicle. Comparing all the state-of-the-art algorithms, approaches based on R-CNN stand out in terms of Average Precision. Typically two-stage detectors have higher accuracy rates while one-stage detectors reaches lower inference times. Mask R-CNN was chosen thanks to the high values obtained for Average Precision without compromising inference times, as well as providing object instance segmentation. This may be useful for approaches such as Multiview in which it is important to match points of the image acquired from one camera with points of the image acquired by other camera in another position. It was necessary to test the adaptability of Mask R-CNN to other datasets by changing COCO dataset to have different number of classes. At the end, the trained network over Bosch dataset, was Faster R-CNN with Mask weights pre-trained over COCO.