Optimized random forest classifier for drone pilot identification

Random forest is a powerful machine learning scheme which finds applications in real-time systems such as unmanned aerial vehicles. In such applications not only the classification performance is relevant but also several non-functional requirements including the classification time, the memory usage and the power consumption. This paper proposes a new approach to improve the real-time behavior of a random forest classifier. This is accomplished by reducing the number of evaluated nodes and branches as well as by reducing the branch length in the underlying binary decision trees with numerical split values. A hardware architecture is presented for the improved tree-based classification method. A proof-of-concept implementation on an FPGA platform and some preliminary results show the advantage of this approach compared to related work.