Object Detection Algorithm Based On Hourglass Module And Its Deployment On Arm

Object detection is an important direction in the field of computer vision.With the development of deep learning methods and related hardware equipment,it has a wide range of applications in autonomous driving,intelligent monitoring,smart phones,etc.,and has important research significance.Anchor-free based object detection is a mainstream one-stage object detection algorithm.An hourglass network structure that incorporates multiple layers of supervisory information can significantly improve the accuracy of the anchor-free object detection algorithm,but its speed is much lower than that of a common network at the same level,and the features of different scale targets will interfere with each other.In order to solve the above problems,an object detection algorithm based on asymmetric hourglass network structure was proposed.The proposed algorithm is not constrained by the shape and size when fusing the features of different network layers,and can quickly and efficiently abstract the semantic information of network.Aiming at the problem of object detection at different scales,this thesis designs a multi-scale output hourglass network structure to solve the problem of mutual interference between features of different targets,and uses a special non-maximum suppression algorithm to refine the output detection results.On the other hand,on embedded and mobile devices,the existing mainstream small detection models are fast but generally unable to achieve high accuracy.Based on the asymmetric hourglass network,this thesis proposes two methods of model pruning and mixed precision training for object detection without anchor,so that the network model can greatly increase its speed while maintaining accuracy.Finally,this thesis migrated model to Arm equipment and successfully deployed it on inspection robot,achieving good application effects.Experimental results show that the accuracy of the algorithm proposed in this thesis has reached a good level on the two mainstream data sets of COCO and VOC,and the two model acceleration methods improve the detection speed and reduce the training resource occupation,achieved both good and fast results.