Perception Maps for the Navigation of a Mobile Robot using Ultrasound Data Fusion

Abstract

Despite appearing insufficient for robotics global navigation, sonar is still meaningful when short-range (local) navigation is concerned. Still, ultrasonic data interpretation is affected by problems such as specular reflections or sensor crosstalk. The first of these problems occurs quite often in navigating situations, and the reliance on a single sensor has little chance of acceptable success. A consequent possibility is using multiple sensors in appropriate geometric lay-outs to overcome this limitation. The paper focuses on the problem of local navigation by using special perception maps built after multi-sensorial ultrasonic data. The main properties of these maps are their topology and geometry. These properties are adapted to sensors characteristics, measurement reliability and spatial redundancy. This results in a non-Cartesian grid where the decision to consider a cell free or occupied comes from more than one sensor. These maps are built using exclusively real data directly from a 24-ultrasonic sensor array. For the effect, a neural network performs the mapping between ultrasonic data and cells’ occupancy. A 3-layer supervised learning network is trained with real data and gives as output the state of each grid cell. A room model is needed during the learning phase to generate the training set, and is no longer required during the operating phase. Large amounts of data are needed so most of the representative situations of navigation are covered in the training set. The networks converged slowly but very efficiently. After the training phase, all the presented patterns matched perfectly. The network is able to cope with most of the specular reflection situations. This is due to the inherent sensor multiplicity and network integrating capabilities. Changes in environment, such as obstacles in vehicle trajectory, are also detected, providing therefore an indication to the network generalisation properties. In a near future, the robot is expected to run autonomously based on these perception maps.