Binocular Human Body Attitude Distance Localization Recognition Algorithm Based on Dual Convolution

Abstract

Objective: An algorithm based on real-time accurate target recognition and distance location for VAVs, the existing target location and discrimination methods often fail to meet practical requirements. Method: The image information collected under the common single camera can only obtain two-dimensional information, and the relative distance of the camera based on the target cannot be obtained. However, the commonly used dual camera-based distance acquisition algorithm is too complicated, not stable enough, and requires developers to have higher the level of knowledge, the high threshold for development, and the difficulty of application. Therefore, this paper proposes to train the feature extraction network based on the two-channel Darknet-53 basic structure through the dual camera under the human body posture recognition image dataset, and initialize the YOLOv2 network with its parameters, and to train the human body position in the human body posture image, relative distance, and category. Result: Experimental results verify that the human body position and category recognition of human posture using this method improves the recognition accuracy by 3.83% and 4.81% compared with the single-convolution chain, and the accuracy of the target-based relative distance is achieved 65.21%. Conclusion: The algorithm can be effectively applied to the UAV to quickly recognize the human body posture and obtain a better recognition effect to meet the real-time demand.