Multi-stage deep learning technique with a cascaded classifier for turn lanes recognition

The accurate recognition of road markings such as lanes and turn arrows is required in many applications including autonomous vehicles. Nevertheless, studies on road markings detection are commonly found in literature, detection and classification of turn lane arrows has not gained much attention. Most of the research which exists on the detection and classification of turn lane arrows have many limitations including low accuracy. Therefore, a novel technique based on two novel concepts for improving the performance of the detection and classification of turn lane arrows is proposed in this paper. Firstly, pixel-wise segmentation of all turn lane arrows into one class instead of each turn lane arrow in a separate class is proposed. Secondly, a novel cascaded classifier that evolves its weights so that it can identify turn lane arrows is proposed. Three turn lane road markings named left turn lane, right turn lane and Continuous Central Turning Lane (CCTL) are evaluated using a real-world roadside image dataset created by video data including all state roads in Queensland provided by our industry partners. The comparative analysis of the experimental results demonstrated outstanding results in terms of accuracy.