A Benchmark Comparison of Visual Place Recognition Techniques for Resource-Constrained Embedded Platforms.

Autonomous navigation has become a widely researched area of expertise over the past few years, gaining a massive following due to its necessity in creating a fully autonomous robotic system. Autonomous navigation is an exceedingly difficult task to accomplish in and of itself. Successful navigation relies heavily on the ability to self-localise oneself within a given environment. Without this awareness of one’s own location, it is impossible to successfully navigate in an autonomous manner. Since its inception Simultaneous Localization and Mapping (SLAM) has become one of the most widely researched areas of autonomous navigation. SLAM focuses on self-localization within a mapped or un-mapped environment, and constructing or updating the map of one’s surroundings. Visual Place Recognition (VPR) is an essential part of any SLAM system. VPR relies on visual cues to determine one’s location within a mapped environment. This thesis presents two main topics within the field of VPR. First, this thesis presents a benchmark analysis of several popular embedded platforms when performing VPR. The presented benchmark analyses six different VPR techniques across three different datasets, and investigates accuracy, CPU usage, memory usage, processing time and power consumption. The benchmark demonstrated a clear relationship between platform architecture and the metrics measured, with platforms of the same architecture achieving comparable accuracy and algorithm efficiency. Additionally, the Raspberry Pi platform was noted as a standout in terms of algorithm efficiency and power consumption. Secondly, this thesis proposes an evaluation framework intended to provide information about a VPR technique’s useability within a real-time application. The approach makes use of the incoming frame rate of an image stream and the VPR frame rate, the rate at which the technique can perform VPR, to determine how efficient VPR techniques would be in a real-time environment. This evaluation framework determined that CoHOG would be the most effective algorithm to be deployed in a real-time environment as it had the best ratio between computation time and accuracy.