Comparison of random forest and long short-term memory network performances in classification tasks using radar

Robust semantic knowledge of the environment is one of the building blocks for autonomous driving. If different sensor types are employed for the same task independently, the overall accuracy and safety of the system can increase. Therefore, it is desirable to maximize each sensor's capabilities and to build up redundancies, as it is often required by functional safety. To this end, this paper demonstrates how classification of dynamic objects using solely radar sensors can be performed. Two different methods are utilized and compared: a random forest classifier and a long short-term memory network (LSTM).