Modelling the Environment of a Mobile Robot using Feature Based Principal Component Analysis

We present a method of classifying the robots environment according to the sensor perception. Our approach combines both the feature based and statistical approaches. We use laser scans at few (nearly five percent) of all the possible poses in a static indoor environment. Then we extract some vital features of lines and corners with attributes such as slope of lines and distance between corners from the raw laser data and compute the associated probabilities of detection. Bootstrap method is used to get a robust correlation of features and finally principal component analysis (PCA) is used to model the environment. In PCA, the underlying assumption is that data is coming from a multivariate normal distribution. The use of bootstrap method makes it possible to use the observations data set which is not necessarily normally distributed. Error and confidence estimates were also established using bootstrap. This technique lifts up the normality assumption and reduces the computational cost further as compared to the PCA techniques based on raw sensor data and can be easily implemented in dynamic indoor environment. The knowledge of the environment can also be updated in an adaptive fashion. Results of experimentation in an office environment under varying environmental conditions using a real-time robotic software Player/Stage are shown