Recognition System of Hand Signals of a Police Officer for Automated Driving.

Current road traffic law prescribes that hand signals performed by a police officer has higher priority compared with that of traffic lights. Therefore, in automated driving system of SAE level 3 or higher, the system needs to recognize the instruction from the motion of the police officer. We developed a method to recognize such hand signals from on-vehicle camera, based on deep-learning technique. The skeleton coordinate of the performer is input to a deep learning method, to classify the signal state into Red/Green or Red/Green/Other. From the state and the continuation conditions, the instruction Stop/Go is determined. Our preliminary experiment proved that quite similar short actions are included both in Red and Green, and it is better to separate such actions as "Other". In the final result, Stop/Go can be appropriately determined, and at the same time, the temporal difference of estimation between switching Stop/Go (Too-early Go and Too-late Stop) was less than 0.43 seconds.