Research Of Multi-task Detection System Based On High-order Fusion Of Convolutional Feature Hierarchical Representation

With the success of deep learning in computer vision,object detection based on convolutional neural network(CNN)is developing rapidly.As one of the hotspots in the field of computer vision,object detection has a wide range of applications in the field of video surveillance,industrial robots,automatic grasping,and so on.In automatic crawling applications,the actual scene usually contains a large number of small target objects,and it needs accurate target pose estimation.The existing CNN based object detection algorithms are usually designed for large target objects,lack of positioning ability for small targets,and can not estimate the target pose changes at the same time.Aiming at the above problems,this paper,starting from the perspective of HyperColumn,introduces the high order feature,which has been used in image classification into object detection.In case of no loss of position information,this paper used two high order feature schemes in detection.One model is second order response transformation(SORT)method based on the characteristics of nonlinear transformation,and another model is high order integration of hierarchical convolution activations(HIHCA)from the view of kernel method.Through the experiment on PASCAL VOC dataset,the second order information used in classification tasks can improve the actual accuracy of object detection in a certain extent.In order to use the characteristics of convolutional neural networks with higher resolution for further,this paper combine TDM model(a feature confusion method)with the HIHCA method,which prove adding higher resolution feature can improve the efficient of detection system to capture the small objects in PASCAL VOC dataset.At the same time,in order to extend the applications of object detection in grasping and vision enhancement,this paper used rigid pose estimation task in object detection system.Through the experiment in PASCAL 3D+,this paper verified the validation of detection with pose estimation based on deep learning.Finally,two models are proposed to form a complete object detection system,and the performance of the proposed algorithm is verified by actual robot target grabbing task.