Enhancing HD-Map Updates Using Iterative Refinement Deep Q-Learning Approach

HD-Map updates are essential for different automotive industry stakeholders. The main parties benefiting from more accurate HD-Map updates are municipalities and automotive manufacturers, especially autonomous vehicle manufacturers. The similarity between the HD-Map update problem to an object detection problem is uncanny. The deep-learning-based object detection algorithms used in the current HD-Map update lack the refinement needed for an efficient learning process. This problem arises from the high dimensionality of the sampling space of the visual input used. In this paper, we propose an iterative refinement process that trains a Deep Q-network to estimate optimal refinement steps for the 2D object detection problem. We verify that on a new efficient synthetic dataset, then we use DQN to enhance the detection accuracy of a multi-class HD-Map change detection model. We benchmark our results on the “Trust but Verify” dataset and achieve an mAP of 82% using a smaller image size and model.