On the efficiency of population-based optimization in finding best parameters for RGB-D visual odometry

Visual odometry estimates the transformations between consecutive frames of a video stream in order to recover the camera’s trajectory. As this approach does not require to build a map of the observed environment, it is fast and simple to implement. In the last decade RGBD cameras proliferated in roboTIcs, being also the sensors of choice for many practical visual odometry systems. Although RGB-D cameras provide readily available depth images, that greatly simplify the frame-to-frame transformations computaTIon, the number of numerical parameters that have to be set properly in a visual odometry system to obtain an accurate trajectory estimate remains high. Whereas seƫng them by hand is certainly possible, it is a tedious try-and-error task. Therefore, in this article we make an assessment of two population-based approaches to parameter opTImizaTIon, that are for long time applied in various areas of robotics, as means to find best parameters of a simple RGB-D visual odometry system. The optimization algorithms investigated here are particle swarm optimization and an evolutionary algorithm variant. We focus on the optimization methods themselves, rather than on the visual odometry algorithm, seeking an efficient procedure to find parameters that minimize the estimated trajectory errors. From the experimental results we draw conclusions as to both the efficiency of the optimization methods, and the role of particular parameters in the visual odometry system.

---

Opracowanie rekordu w ramach umowy 509/P-DUN/2018 ze środków MNiSW przeznaczonych na działalność upowszechniającą naukę (2019).