Classification of Degraded Traffic Signs Using Flexible Mixture Model and Transfer Learning

Automatic detection and recognition of traffic signs is a topic of research for various applications like driver assistance, inventory management and autonomous driving. Poorly maintained traffic signs degrade by losing their colors or some part is weird due to aging and hence making the task more challenging. The problem is mainly related to the developing world and has gained less attention in the literature on automatic traffic sign detection and recognition. To handle the degradation issue, we present a novel flexible Gaussian mixture model based technique with automatic split and merge strategy. This adaptive scheme works as a preprocessing step which facilitates locating traffic signs under a certain degree of degradation in a real world scenario. A multiscale convolutional neural network augmented with dimensionality reduction layer is proposed to recognize contents of the sign. Since, there is no available benchmark dataset for this purpose, we collected a number of images containing partially degraded signs from the famous German Traffic Sign Detection Benchmark (GTSDB) and augmented it with manually and naturally degraded traffic sign images taken from the longest highway in the country of authors' residence. Experimental results show that our proposed technique outperforms many state of the art and recent methods for detection and recognition of degraded traffic signs.