Neural Network-POMDP-Based Traffic Sign Classification Under Weather Conditions

Abstract

Despite their initial success in operating autonomously, self-driving cars are still unable to navigate under severe weather conditions. In the proposed system, a multi-layer perceptron (MLP) performs initial traffic sign classification. The classification is input as an observation into a partially observable Markov decision process (POMDP) in order to determine whether taking another picture of the sign, or accepting the classification determined by the MLP is the more optimal action. The synergistic combination of the MLP with the POMDP was shown to have a greater functionality than the sum of the MLP and POMDP operating in isolation. The results demonstrate the MLP-POMDP system is capable of training faster and more accurately classifying traffic sign images obscured by fog than a MLP. With further development of this model, one of the greatest shortcomings of autonomous driving may be overcome by accurately classifying signs despite obstruction by weather.