Improving Visual Place Recognition in Changing Environments

For many years, the research community has been highly interested in autonomous robotics and its various applications, from healthcare to manufacturing, transportation to construction, and more. An autonomous robot's key challenge is the ability to determine its location. A fundamental research topic in localization is Visual Place Recognition (VPR), a task of detecting a previously visited location through visual input alone. One specific challenge in VPR is dealing with a place's appearance variation across different visits, which can occur due to viewpoint and environmental changes such as illumination, weather, and seasonal variations. While appearance changes already make VPR challenging, a further difficulty is posed by the resource constraints of many robots employed in real-world applications that limit the usability of learning-based techniques, which enable state-of-the-art performance but are computationally expensive. This thesis aims to combine the need for accurate place recognition in changing environments with low resource usage. The work presented here explores different approaches, from local image feature descriptors to Binary Neural Networks (BNN), to improve the computational and energy efficiency of VPR. The best BNN-based VPR descriptor obtained runs up to one order of magnitude faster than many CNN-based and hand-crafted approaches while maintaining comparable performance and expending a small amount of energy to process an image. Specifically, the proposed BNN can process an image 7 to 14 times faster than AlexNet, spending 13\% of the power at most when deployed on a low-end ARM platform. The results in this manuscript are presented using a new performance metric and an evaluation framework designed explicitly for VPR applications aiming at the two-fold purpose of providing meaningful insights into VPR performance and making results easily comparable across the chapters.