One of the major application scenarios of military and police unmanned vehicles is to support and carry equipment to reduce the load of personnel. Therefore, personnel identification technology in complex off-road environment is one of the key technologies of military unmanned vehicles. The effect of laser radar identification is relatively good, but the method of monocular vision can greatly reduce the system cost and facilitate the popularization and application of equipment. Aiming at the problem of personnel identification under monocular vision perception, especially the problem of background interference in complex environment, this paper designs an improved YOLOV3 deep learning network architecture based on monocular vision information, and proposes an improved YOLOv3 deep learning network architecture of residual module. in the original YOLOv3 deep learning network, the signal processed by two CBL modules on the input of residual module RB is added to the input of RB module. The difference between the improved YOLOv3 deep learning network structure of residual module and the original YOLOv3 deep learning network structure is mainly reflected in the design of residual module, and the result is the output of RB module. In the YOLOv3 deep learning network with improved residual module, a global pooling layer, two full connection layers, Sigmoid activation function layer and multiplication layer are added to the original YOLOv3 network RB output. Through the verification of real vehicles in off-road environment, the results show that the designed personnel recognition algorithm based on monocular vision can accurately recognize personnel targets in complex environment, and the recognition rate in complex environment exceeds 95%.

CICTP