Research On The Theories And Applications Of Object Detection Algorithms In Complex Scenes

As one of the classic topics within the field of computer vision,object detection features a wide selection of applications in many fields,such as autonomous driving,robot vision,medical image analysis,human-computer interaction and so on.Especially in recent years,with the speedy development of deep learning technology,object detection technology has created breakthrough progress.However,thanks to the complexness and variability of the particular scene,there will be occlusion,density,scale change,illumination and other factors,which may adversely have an effect on the performance of the object detection algorithm.And how to effectively solve the matter of object detection in complex scenes has been one in every of the analysis difficulties.Therefore,this paper studies the theories and applications of object detection algorithm in complex scenes.The research during this paper is as follows:(1)In this paper,a mask wearing detection algorithm called CSPC-YOLO in complex scenes is proposed,which combines cross-stage partial network and CIo U.Firstly,the Dark Net53 backbone network is improved based on the cross-stage partial network,which reduces the calculation consumption and increases the training speed.Secondly,an improved spatial pyramid pooling structure is introduced into YOLOv3,and the top-down and bottom-up feature fusion strategies are used to optimize the multi-scale prediction network,so as to realize feature enhancement.Thirdly,CIo U is selected as the loss function,which fully takes into account the information such as the central point distance,overlap area and aspect ratio of target box and predicted box.Finally,experiments are carried out on the newly marked mask wearing detection data set.The experimental results show that the new method can effectively improve the accuracy and speed of mask wearing detection task in complex scenes,with an average accuracy of90.2% and a detection speed of 38 FPS.(2)In this paper,a helmet wearing detection algorithm called ACGCSPC-YOLO in complex scenes is also proposed,which combines asymmetric convolution network and attention mechanism.Compared with the mask wearing detection data set,the helmet wearing detection data set adds complex scenes such as noise and color gamut distortion,which increases the difficulty of object detection.Therefore,the ACGCSPC-YOLO algorithm is improved based on the CSPC-YOLO algorithm.Firstly,the Dark Net53 backbone network is further improved based on the asymmetric convolution network,which enhances the feature extraction capability of the backbone network.And the network computation is reduced by using 1 × n and n × 1 serial convolution instead of n× n convolution in some convolution blocks.Secondly,the multi-scale prediction network is further improved by embedding CBAM attention mechanism to improve the detection accuracy.Thirdly,the loss function is further optimized by combining Gaussian model.Finally,experiments are carried out on the newly helmet wearing detection data set.The experimental results show that the new method can effectively improve the accuracy and speed of helmet wearing detection task in complex scenes,with an average accuracy of94.6% and a detection speed of 57 FPS.