A foreground-focused action recognition algorithm for intelligent unmanned systems

Abstract

A foreground-focused action recognition algorithm based on convolutional neural network for intelligent unmanned systems is presented in this paper. The algorithm consists of the foreground extraction part and the action recognition part. For the foreground extraction part which can be considered as an object detection task, a convolutional neural network-based single-shot model with the spatial pyramid pooling structure and Mish activation function is developed to achieve a better performance. For the action recognition part, a Convolution-3D modified model is proposed which applies the 3D convolution kernel to obtain temporospatial features. To avoid the distortion of the foreground target, the padding form of the foreground extracted by the object detector is used as an input of the action recognition part. The algorithm can be applied to intelligent unmanned systems, being able to make a real-time performance of accurate action recognition.