Aiming at real-time gesture guidance and automatic transfer of shipborne UAV,the dynamic gesture recognition method based on convolutional neural network has some problems that need to be solved urgently,such as boundary indistinction of start position of gestures and many times responses of a gesture.Therefore,an online dynamic gesture recognition algorithm based on single-model architecture is proposed.The lightweight(2+1)D conv-olutional neural network is used to extract features of dynamic gestures.The class score vector which is output by net-work is added to the fixed length cache queue,and the average vector of class score is calculated.Then the start posi-tion of gestures is detected through threshold judgment,and the occasional false detection of the recognition network is filtered out.When gesture start is detected,the average vector of class score is added to the decision queue,and the output of gesture recognition result is controlled by a flag to achieve the purpose of single response.The experimental results show that the algorithm can detect the start position of gestures and solve the problem of many times responses of a gesture.The Levenshtein accuracy of 86.91%is achieved on the self-built shipborne UAV transfer gesture data-set,and the speed of real-time detection is 60 frames per second.
关键词
舰载无人机转运/动态手势识别/实时/手势起始位置检测/单次响应
Key words
Shipborne UAV transfer/Dynamic gesture recognition/Real time/Detection of start position of gestures/Single response
维普
万方数据