Traffic sign recognition on Indian database using wavelet descriptors and convolutional neural network ensemble

Abstract

Traffic sign recognition (TSR) has been a rising and lucrative field for researchers during the last decades. The high improvement of ADAS (autonomic driving autonomous system) has led researchers worldwide to concentrate on the development of TSR systems. As such, a novel automated multiclass TSR system is proposed. The architecture uses wavelet descriptors to extract the high information density and the traffic signs' edges and curves. The LL band image is directly fed into the classifier to avoid normalization. Three classifiers, CNN, CNN ensemble, and LSTM, are deployed for recognition. The architecture is implemented on IRSDBv1.0, the first available Indian traffic sign database. The architecture is also implemented on the standard traffic sign database GTSRB to investigate its effectiveness. An efficiency of 71.57% and 96.76% are recorded on IRSDBv1.0, and GTSRB, respectively. A list of comparative results is also provided to prove the competence of the architecture. The reasons behind the difference in the achieved accuracy are also discussed.