| Object detection is one of the main tasks of computer vision.The goal is to give the category of the object in the image and the position coordinates of the object.In recent years,with the rapid development of deep learning,the use of convolutional neural networks for object detection has become a trend,and object detection based on convolutional neural networks can achieve higher detection accuracy and faster detection speed.However,object detection is still a challenging task due to problems such as multi-scale and occlusion.Two problems are mainly studied in this paper,one is multi-scale object detection,the other is small and occluded object detection.In order to solve the problem of multi-scale object detection,three multi-scale object detection methods are proposed in this paper.(1)The feature pyramid object detection method based on VGG16 is implemented in this paper,then three additional downsampling layers are added,and an improved feature pyramid object detection method is proposed.(2)Then the concatenation of convolutional layers is studied and finally a feature pyramid method with intra-stage aggregation is proposed.(3)Finally,the intra-stage aggregation method and feature pyramid method are combined in this paper,and in order to improve the training speed,the normalization layer is added.Finally,an optimized multi-scale object detection method is proposed.To solve the problem of small object and occluded object,the context information module is studied in this paper,and two context information modules are proposed.(1)Multi-branch context module:this module enhances the receptive field of the feature by using convolution branches of three different convolution kernel sizes of 1x1,3x3,and 5x5.(2)ROI context module:this module obtains the context area by expanding the object candidate box to assist in the classification and regression of the second stage.In this paper,all the proposed methods are tested on KITTI dataset,and some of them are tested on PASCAL VOC0712 dataset.Experiments show that the proposed method can improve the accuracy of object detection,and eventually achieve 88.19%detection accuracy on KITTI Moderate dataset,and 0.7%improvement on PASCAL VOC 0712. |
