Vibration-Based Terrain Classification for an Autonomous Truck

Abstract: This thesis is focused on developing vibration based terrain classification for an autonomous mining truck. The goal is to classify between good and bad gravel roads as well as good and bad asphalt roads. Current literature within vibration based terrain classification has been focused to a great extent on smaller research vehicles. On smaller research vehicles have roll-rate, pitch-rate and vertical acceleration been reported to yield the highest average classification rates. Common approaches for pre-processing the data consists of segmenting the data, apply filtering techniques, computing the Power Spectra Density (PSD), performing Principal Component Analysis (PCA) and compute the logarithms. How to do this specifically for an Autonomous Truck (AT) is not trivial. What signals from the trucks Internal Measurement Unit (IMU)s yields the highest average classification rates? How does one process the raw data in the best way, and what classification method performs the best for this for an AT? The AT studied here have five different IMUs that all measure ẍ, ÿ, z̈ acceleration, and ωroll, ωpitch, ωyaw rotational speed. One is located in the cab, and the other four are located in each of the four corners of the chassis. With these sensors empirical vibration data from different surfaces, speeds and loads was gathered with multiple identically equipped autonomous mining trucks. With this data were experiments conducted in order to find a high performing classifier that also was possible to implement in the ATs software in C++. The different signals were ranked according to the highest classification score, and different pre-processing parameters combined with different classification methods likewise were. ωyaw and ωpitch from the cab IMU, and z̈ from the rear right IMU were the ones that yielded the highest average classification rates. The pre-processing consists of segmenting the data, multiplying the segment with a window function, compute the one-sided PSD, logarithmize the PSD values and lastly normalize the data. A bagged classifier based on Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel showed the highest classification performance. The final multiclass classifier was a combination of three of these bagged classifiers in a tree structure. The F-measure rates for the four classes were {0.946, 0.98, 0.714, 0.879}.