Fast Multi-Target Recognition Method for Humanoid Robot Playing Soccer

Fully autonomous humanoid robot soccer competition requires the dynamic-featured robot recognizing multi-target both accurately and rapidly under complex field environment.A fast multi-target recognition method for humanoid soccer robot is proposed.Haar-like features integrated with morphological analysis is employed to construct heat maps,on which the candidate patches supposed containing object in the robot’s field of view are roughly selected to form a set of“region of interest(ROI)”.This ROI set is quickly classified into target or non-target regions through a lightweight Tiny-DNN convolution neural network.Meanwhile,each ROI generated automatically during the robot operation can be directly collected and stacked into positive or negative sample sets for the off-line training of the convolution neural network,thus avoiding tremendous labor work and uncertainty deviation of manually tagger of target sample on quantities of pictures.The proposed method is applied to CPU supported SYCU-Legendary humanoid robot,enabling it to recognize 95.8%,96.2%and 96.0%of football,goalposts and penalty marks within 0.03 seconds.Besides,foreign matters and light changes cast little influence on targets recognition so that SYCU-Legendary team won the championship in 2018 and 2019 RoboCup China Open,which indicates the method in this paper is worthy to be widely extended to other application domains.