Recognition And Tracking Of Occluded Targets In Complex Indoor Scenes

Target recognition and tracking is one of the hotspots for computer vision,and it is also an important generic technology for service robots.The occlusion of targets could lead to some problems,such as feature points missing and incomplete contours.It has become a huge problem for target recognition and tracking in complex scenes,and it is also a common problem in machine vision applications.Recognition is of great significance for target search,hand-eye coordination,and security patrol with service robots.So far,there are still many problems in this field,this thesis mainly focuses on target detection in occluded scenes and occluded target tracking with dynamic vision of service robots.The research contents are as follows:(1)In order to reduce the false recognition rate caused by occlusion in indoor complex scenes,a detection algorithm based on multi-source feature information fusion is proposed.Firstly,the particularity of the single feature in multi-source features is used to identify and locate the occlusion region in images.And then,fusion algorithm is used to identify the filtered position.This method reduces the calculation of consumption for feature information by cutting down the detection area.The experimental results show that this algorithm improves the real-time performance and accuracy for target recognition.(2)A detection method named YOLO-V3-Efficient is proposed to solve the difficulty of key features extraction from the feature extraction layers caused by different resolution and deformation of input images.This detection method combines the feature extraction layers of EfficientNet and YOLO-V3 network.Extracted features from each layer and the resolution of inputs becomes a scaled YOLOV3?baseline model.The improved model can adjust the depth and width of the convolutional layers and the resolution of the input image,meanwhile,it can solve more feature omission problems caused by image deformation and occlusion.Experiments show that the performance index of the proposed method such as mAP,FPS and so on has advantages over the YOLO-V3,Fast-RCNN and SSD.A higher accuracy and real-time performance have been achieved.(3)An ATLD tracking algorithm based on adaptive update template is proposed to solve the problem of variable scene and target occlusion when the camera is moving.This method needs to set an occlusion decision condition first.When occlusion is triggered,the original tracking algorithm is used to obtain the target information before occlusion,and then the template is replaced.The position of the target is updated by motion estimation and compensation,and the position of the target at the next moment is predicted by Kalman filter.The experimental results show that the ATLD algorithm can track the target in real time.