Fixed-Point Neural Network Ensembles for Visual Navigation

Visual navigation is an important research field in robotics because of the low cost and the high performance that is usually achieved by visual navigation systems. Pixel classification as a road pixel or a non-road pixel is a task that can be well performed by Artificial Neural Networks. In the case of real-time instances of the image classification problem, as when applied to autonomous vehicles navigation, it is interesting to achieve the best possible execution time. Hardware implementations of these systems can achieve fast execution times but the floating-point implementation of Neural Networks are commonly complex and resource intensive. This work presents the implementation and analysis of a fixed-point Neural Network Ensemble for image classification. The system is composed by six fixed-point Neural Networks verified with cross-validation technique, using some proposed voting schemes and analyzed considering the execution time, precision, memory consumption and accuracy for hardware implementation. The results show that the fixed-point implementation is faster, consumes less memory and has an acceptable precision compared to the floating-point implementation. This fact suggests that the fixed point implementation should be used in systems that need a fast execution time. Some questions about ensembles and voting have to be reviewed for fixed-point Neural Network Ensembles.