Q. Zhou, S.F. Wang, Y.Q. Wang, J.Y. Zhang

Pages: 159-170

Abstract
For autonomous vehicles, when encountering traffic accidents or bad weather and other complex situations that made the traffic signal stop working, recognizing the command of traffic police gesture was very important. A new method of traffic police gesture recognition was proposed. First, the traffic police gestures were decomposed into "key actions" and "transition actions" based on the motion characteristics, Secondly, using Faster R-CNN and fuzzy matching the gesture recognition model was build. Faster R-CNN was used to extract the features of decomposed actions for image recognition; a fuzzy matching method was built to match gestures based on the recognition of decomposed actions. Finally, the recognition experiment was executed, as for common traffic police gesture, the result showed that the new proposed method was effective.
Keywords: traffic police gestures; image recognition; decomposition action; Faster R-CNN; fuzzy matching