Extended Single Shoot Multibox Detector for Traffic Signs Detection and Recognition in Real-time

Real-time traffic signs detection and recognition is an essential task in autonomous driving and technology-assisted driving. In the past decades, a significant improvement in traffic sign detection and recognition has been gained due to deep learning methods. However, there is still considerable room for improvement. The traffic sign appearance variations limit the performance of the current state of the arts from perfection. Previous results have shown that the detection performance degrades for low-resolution and far scale traffic signs, resulting in high localization and recognition error and degrades CNN's object representation. Since feature maps of top layers of CNN illustrate low-level patterns that cause learning and discrimination of feature patterns challenging for the downstream recognizer. Significantly when a small-scale object passes through the layers likely to vanishes in the middle of the framework. To compensate for the loss of low-level details and localization, we proposed an Extended Single Shoot MultiBox Detector(ESSD), which integrates high resolution and easy-to-compute handcrafted feature channels with the low resolution computational expensive CNN feature channels. We apply DeepMultiBox on the hand-engineered channels (i.e., HOG+LUV and Fishers Discriminant Analysis HOG(FDA HOG) and CNN feature maps in parallel to keep a full set of object representation. We also apply color channels and shape information in a novel way to refine the localization. Experimental results on the German TSR(Trafic signs recognition) benchmark dataset shows the efficiency of the proposed approach, which is 99% mAP@0.5 IoU (intersection over union) and processes up to 60+ FPS(frames per second) in real-time and achieves recognition accuracy competitive to the state of the arts.