Improved kernel DM+V with adaptive Bi-variate distribution model and distance transform for gas distribution mapping using mobile robot

Mobile robot olfaction is a complex research area, which can be broken down into several major challenges; one which is Gas Distribution Mapping (GDM). GDM is a representation of how gas is spatially dispersed within an environment. Additionally, GDM can also be used to localise the source of leaking gas. Gas dispersion is affected by various factors such as airflow, gas concentration gradient and obstacles. Current research in GDM has only considered airflow effects and gas concentration gradient but has so far neglected obstacles. Hence, this research proposes a new improvement on the Kernel DM+V, a statistical-based GDM method, to take into account walls of the environment in estimating GDM; and consequently, improve the gas source localisation in indoor environment. Preliminary studies in this research established that the usable kernel width for Kernel DM+V is in the range of 0.5m ≤ σ ≤ 2.5m. Further GDM experiments showed that better likelihood, estimated through Negative Log Predictive Density (NLPD), in GDM were found at higher kernel widths. On the other hand, gas source localisation accuracy was also analyzed and results show that larger kernel width worsens the accuracy of gas source localisation. Hence, two solutions were proposed to improve GDM and gas source localisation by considering the walls in the environment. Firstly, a known map obtained through Simultaneous Localisation and Mapping (SLAM) was used to compute Distance Transform (DT) of the environment to obtain a better correlation among cells in the grid map. The proposed solution was aimed to improve the weighing function of the Kernel DM+V based on the structure of the environment. Next, the Kernel DM+V was presented as a Bi-variate distribution model to extend the method to unknown environments. By using a potential field estimated based on walls and free-space in the environment, the extrapolation kernel of Kernel DM+V is modified to suit the structure of the environment. This method constantly changes the kernel width based on the structure of the environment as the mobile robot explores the environment. Both GDM and gas source localisation was demonstrated to be improved compared to the conventional method when walls information is taken into consideration in estimating GDM. The research has established that the method can be incorporated into existing GDM methods to improve gas distribution estimation in indoor environments where walls are prevalent.