Visual Place Recognition in Changing Environments Utilising Sequence-Based Filtering and Extremely JPEG Compressed Images

Visual Place Recognition (VPR), part of Simultaneous Localisation and Mapping (SLAM), is an essential task for the localisation process, where each robotic platform is required to successfully navigate through its environment using visual information gathered from the on-board camera. Despite the recent efforts of the research community, VPR remains an improving process. To this end, a large number of deep-learning-based and handcrafted VPR techniques (also referred as learnt and non-learnt VPR techniques) have been proposed to overcome the challenges in this field, such as viewpoint, illumination and seasonal variations. While Convolutional Neural Network (CNN)-based VPR techniques have significant computational requirements that may restrict their applicability on resource-constrained platforms, handcrafted VPR techniques struggle with appearance changes. In this thesis, two mainly unexplored avenues of research are investigated, namely sequence-based filtering and JPEG compression. To overcome the previously mentioned challenges, this thesis proposes a handcrafted VPR technique based on HOG descriptors, paired with an adaptive sequence-based filtering schema to perform VPR in scenarios where the appearance of the environment drastically changes upon different traversals. The technique entitled ConvSequential-SLAM is capable of achieving comparable place matching performance with state-of-the-art VPR techniques at reduced computational costs. The approach utilised for matching sequences of images in the above technique has been employed to investigate the improvement in VPR performance and the computational effort required to execute VPR when utilising a sequence-based filtering approach. As CNNs are computationally demanding, this thesis shows that VPR can be performed more efficiently using lightweight techniques. Furthermore, this thesis also investigates the effects of JPEG compression for VPR applications, where important reductions in both transmission and storage requirements can be achieved. As the VPR performance is drastically reduced, especially for high compression ratios, this thesis shows how a fine-tuned CNN can achieve more consistent VPR performance on highly JPEG compressed data (i.e. above 90% JPEG compression). Sequence-based filtering is introduced to overcome the performance loss due to JPEG compression. This thesis shows that the size of a JPEG compressed image is often smaller than the size of the image descriptor, and therefore should be transferred instead. Furthermore, our experiments also show that the amount of data required for transfer is reduced with an increase in JPEG compression, even when requiring an increased number of images in a sequence. This thesis also analyses the effects of image resolution on the performance of handcrafted techniques, to enable efficient deployment of VPR solutions on commercial products. The analysis performed in this thesis confirms that local feature descriptors are unable to operate on low-resolution images, as no keypoints (salient information) are detected. Moreover, this thesis also shows that the time required to perform VPR is reduced with a decrease in image resolution.