Abstract: As most existing methods for target localization and identification of UAVs are not practical, a new algorithm for real-time accurate target identification as well as distance localization is proposed. In general, commonly used single cameras can only provide two-dimensional information and cannot be used to compute the relative distance between the camera and the target. Distance collection algorithms using dual cameras are often too complex, technically demanding and non-stable, posing high technical threshold and facing difficulties in application. Therefore, this study trains a feature extraction network based on the basic structure of the dual-channel Darknet-53 through human pose recognition dataset with dual cameras, and applies its parameters to initialize the YOLO-V2 network, which is used to detect the position, relative distance, and type of human bodies from human pose images through training. Experiments show that the new algorithm is 3.85% and 4.83% higher in recognition accuracy compared to single convolutional chain, and achieves an accuracy of 65% in target-based relative distance recognition. The algorithm can be effectively used for UAVs to quickly recognize human postures at a long distance and achieve better recognition results to meet real-time requirements.

Key words: convolutionalneuralnetwork, doubleconvolutionchain, binocularstereovision, humanposerecognition