Efficient visual place recognition in changing environments for resource-constrained platforms

Visual Place Recognition (VPR) enables autonomous systems to localize themselves within their environment using image information, a crucial component of robotic navigation. The affordability, availability, and versatility of current camera sensors makes VPR an attractive option for many mobile robotic applications. However, performing highly reliable VPR is a complicated task, as a place’s appearance can be significantly altered with changes in illumination, visiting viewpoint, seasons, weather, and movement of dynamic elements. Conversely, two distant places within the same environment can appear similar, a problem known as perceptual aliasing. In recent years, VPR techniques built upon Convolutional Neural Networks (CNNs) have consistently improved state-of-the-art performance. However, these approaches continuously tend to larger models, increasing computation and thus requiring more powerful hardware. While the training process of these models can take place in extremely powerful hardware, targeted mobile robotic applications often operate with low-end hardware, usually due to cost or size, rendering CNN based solutions unsuitable, and making lightweight VPR techniques highly desirable. This thesis addresses the computational shortcomings of CNN-based methods by proposing novel algorithms to perform highly efficient VPR, primarily focusing on bio- inspired and multi-model approaches. The first two contributed algorithms are based on DrosoNet, an exceptionally compact model inspired by the odor processing abilities of the fruit fly. Using multiple of these small units in tandem, the proposed Region-DrosoNet VPR algorithm achieves comparable performance to expensive state-of-the-art techniques, with an AUC of 0.94 on the challenging Nordland Winter dataset, at a fraction of the computational cost, requiring only 9 milliseconds to perform a match. Moreover, in an attempt to improve efficiency in standard multi-model VPR algorithms, an adaptive technique-switching mechanism is proposed which selects the most apt VPR techniques for the current environment, without access to ground-truth information. The system, dubbed A-MuSIC, achieves an average AUC of 0.88, a substantial increase from the 0.66 of the current multi-technique state-of-the-art, while requiring only 668 milliseconds to perform a match versus the state-of-the-art’s 1698 milliseconds.