基于深度学习的陪护机器人目标检测方法OACSTPCD

In view of the false detection and missed detection when escort robots carry out object detection tasks in complex environments,an escort robot object detection method based on deep learning is proposed.The improved WGAN network is used to generate high-quality images to enrich the diversity of data sets and lay a foundation for subsequent YOLOv5s training.The Gridmask data enhancement technology is used for the original YOLOv5s model to improve the recognition ability of the detection model for complex scenes and occlusion.In the part of Backbone,the module C2fSE is adopted,and LSTM network and YOLOv5s network are combined to reduce the number of network parameters,obtain more context information,and improve the accuracy of object detection.The dice coefficient(DC)loss is added to the total loss function.In comparison with the original algorithm,the detection accuracy,recall rate and mAP of the improved algorithm increases by 6.8%,4.6%and 8.3%,respectively.The experimental results show that the improved algorithm has been improved in the detection accuracy and missed detection of small objects and partially-occluded objects,which is suitable for the object detection by household escort robot.