Quantitative comparison of supervised algorithms and feature sets for traffic sign recognition

Nowadays a timely detection of relevant events and an efficient recognition of objects in an environment is a critical activity for many Cyber-Physical Systems (CPS), which may have severe impact on citizens, infrastructures or the environment when incurring malfunctions. In the automotive domain, the detection and recognition of traffic signs (TSDR) from images was and is currently being investigated as it heavily impacts the behavior of (semi-)autonomous vehicles. Despite many classifiers and feature extraction strategies applied to images sampled by webcams installed on cars have been developed throughout the years, those efforts did not escalate into a clear benchmark nor comparison of the most common techniques on multiple datasets and feature sets. This study tackles this problem by providing a comprehensive quantitative comparison of traditional supervised Machine Learning algorithms with different feature sets and Deep Learning models for the recognition of traffic signs from three publicly available datasets. A perfect classification was achieved by many ML algorithms on the German GTRSB dataset, while with the BelgiumTSC and DITS datasets no algorithms provided perfect classification.