Online terrain classification for mobile robots using FPGAs

This work proposes a Field Programmable Gate Array implementation of a multilayer perceptron neural network for terrain classification. A 3-axis accelerometer was used for acquiring the acceleration variation that a robot suffers when moving on four different terrains: sand, asphalt, grass and soil. A multilayer perceptron neural network was trained in order to perform the classification process. Afterward, the trained weights and biases were used to implement in hardware the mathematical model of the proposed network. The implemented circuits were characterized in terms of the hardware resources consumption, operational frequency and power consumption. Numerical comparisons between hardware and software results were used in order to validate the hardware implementation and to estimate the classification error. In addition, an execution time comparison using three software-based embedded platforms was performed.