Abstract:

Therefore， with the human pose recognition problem of robots as the core， a local feature recognition algorithm for human pose based on an improved RBF neural network is proposed to improve recognition accuracy. Using a depth camera to obtain three-dimensional orientation data of human joint points， normalizing the orientation data， and constructing three-dimensional coordinates of joint points； Considering the differences between different individuals， in order to achieve nonlinear mapping and optimization of human pose data， accurately identify different individual poses， a Newrbe function is used to construct an RBF neural network， extract feature vectors of human pose data， and provide important basis for recognition； To enhance the ability of RBF neural networks to handle different individual pose differences， ensure recognition accuracy and adaptability， particle swarm optimization algorithm is used to improve the neural network， and genetic operations are performed on particles with specific probabilities to achieve network optimization and obtain local feature recognition results of human pose. The experimental results show that the proposed algorithms have relatively low relative errors， can be maintained below 0.8， high recognition accuracy， and the loss function is minimized when the number of iterations reaches 20. The convergence speed is fast， which can provide a solid foundation for human-machine interaction in the field of agricultural mechanization.

Key words: improve rbf neural network, human posture, local feature recognition, three dimensional coordinates, particle swarm optimization