2020/10/03 by Satyajit Tourani, Dhagash Desai, Tourani, Satyajit +11
Engineering · Computer Science · #Robotics and Sensor-Based Localization #Advanced Image and Video Retrieval Techniques #Indoor and Outdoor Localization Technologies
paper · pdf · doi:10.48550/arxiv.2010.01421
Significant advances have been made recently in Visual Place Recognition\n(VPR), feature correspondence, and localization due to the proliferation of\ndeep-learning-based methods. However, existing approaches tend to address,\npartially or fully, only one of two key challenges: viewpoint change and\nperceptual aliasing. In this paper, we present novel research that\nsimultaneously addresses both challenges by combining deep-learned features\nwith geometric transformations based on reasonable domain assumptions about\nnavigation on a ground-plane, whilst also removing the requirement for\nspecialized hardware setup (e.g. lighting, downwards facing cameras). In\nparticular, our integration of VPR with SLAM by leveraging the robustness of\ndeep-learned features and our homography-based extreme viewpoint invariance\nsignificantly boosts the performance of VPR, feature correspondence, and pose\ngraph submodules of the SLAM pipeline. For the first time, we demonstrate a\nlocalization system capable of state-of-the-art performance despite perceptual\naliasing and extreme 180-degree-rotated viewpoint change in a range of\nreal-world and simulated experiments. Our system is able to achieve early loop\nclosures that prevent significant drifts in SLAM trajectories. We also compare\nextensively several deep architectures for VPR and descriptor matching. We also\nshow that superior place recognition and descriptor matching across opposite\nviews results in a similar performance gain in back-end pose graph\noptimization.\n