| Diving is a dangerous activity.With the continuous development of science and technology,human activities underwater are more and more diverse,and divers in deeper and more complex environments also bring more challenges.Underwater machines have replaced divers in many fields.The emergence of underwater robots has largely replaced manual work,but for many complex work and tourist diving,the involvement of divers is still required,which can combine the advantages of both.Gestures are the main means of communication between divers and are also an effective way for divers to issue instructions to underwater vehicles.This thesis studies diver tracking and gesture recognition based on machine vision against the background of underwater human-machine interaction.Tracking and receiving instructions for divers are functions that must be implemented by the cooperative underwater vehicle.This thesis implements the detection of diver and divers’ gestures based on YOLOv5,and implements the multi-target tracking of divers based on DeepSORT using the detection network.Firstly,the general framework of target detection algorithm based on deep learning is studied,the network structure of YOLOv5 is analyzed,and the diver and diver gesture recognition based on YOLOv5 is implemented.By collecting web data,a data set of divers and gestures was created,comprising 17 scenes and a total of pictures.Considering the common color distortion of underwater images,several data enhancement methods are used to improve the robustness of the model and make the network suitable for many complex environments.The difficulties of diver and gesture detection are analyzed experimentally.Secondly,several improvements are made to overcome the shortcomings of YOLOv5 in diver and gesture detection.First,the attention mechanisms of SE,CBAM and CA commonly used in target detection are compared.CBAM attention module is added to YOLOv5 network,which improves the channel and spatial connection of each target feature,and is more conducive to improving the effective features of the extracted target.Secondly,the feature fusion network of YOLOv5 is improved by using transpose convolution combined with BiFPN structure,which enhances the utilization of shallow information of backbone network and makes featu refusion more efficient.Thirdly,the Border Regression Loss Function is improved to make the detection network learn the shape characteristics of the detected objects better.Finally,the non-maximal suppression algorithm is improved and label smoothing is added to make the network model better able to cope with occlusion and over-fitting.Finally,a multi-target tracking algorithm is designed to track and distinguish divers.The multi-target detection algorithm is analyzed according to the task requirements.Using the detection network of divers and gestures,the improved DeepSORT algorithm is used for diver trac king.Kalman filter is used to estimate the motion state information of the tracking target,OSN et is used to extract the appearance information of the target,and cascade matching is used to correlate the target with the track.Finally,the algorithm is verified. |
