Traffic Police Gesture Recognition using RGB-D and Faster R-CNN

How to make self-driving cars understand the traffic gestures of traffic police is crucial for driverless, especially in China there are many police to help the traffic move smoothly and quickly at different intersection in rush hours. Faster R-CNN in deep learning is a mainstream method, however, has a low recognition rate in the case of complex backgrounds. In order to improve the recognition accuracy under complex environment, a two-stream Faster R-CNN based on color and depth data is proposed in this paper. Depth channel information is used to combine with RGB channel information at the feature level. RGB channel information is integrated with Depth channel information based on Faster R-CNN and RGB-D. Experimental results show that this method is more advantageous than the Faster R-CNN using only RGB data.