Object Detection In Complex Scenes

Complex scenes mainly refer to a series of special scenarios that it is difficult to guarantee the accuracy of the algorithm and the speed of the algorithm if continue to use the original detection algorithm.The reasons for the deterioration of the detection algorithm mainly include the low detection accuracy due to the high density,different size,high similarity and easy deformation of the target;and the real-time detection of the target detection due to the insufficiency of the detection process design.How to perform target detection efficiently and accurately according to requirements in complex scenarios has become a hot issue in computer vision.This paper focuses on the problem of object detection in complex scenarios.The specific work and innovations are as follows:(1)The basic theory of convolutional neural networks is analyzed in detail,and the traditional object detection model and the classic model of object detection based on deep learning are summarized to provide reference for the design of innovative models in complex scenarios.(2)Because the non-rigid body is easy to deform,the main difficulty of non-rigid body detection in complex scenes is that the objects are occluded from each other,and the contours of each object are diverse.This paper proposes Distance Adaptive Convolutional Neural Network(DACNN),improves the regression object and loss function,and designs the distance adaptive convolution layer.Experiments show that the model not only has high detection accuracy,but also has fast detection speed and good robustness,and has broad application prospects.(3)Due to the high degree of similarity between the objects of the rigid body,the main difficulty in the detection of rigid bodies in complex scenes is the easy misclassification caused by the high similarity of appearance(including color,texture,shape,etc.)between the objects.This paper proposes Rigid Region-based Convolutional Neural Network(Rigid R-CNN),a simple and efficient detection network model,improved loss function and non-maximum value suppression algorithm to improve detection accuracy,simplify network structure and improve feature pyramid network to improve detection speed.Experiments show that the model has high detection accuracy for rigid body detection in complex scenes,and has good robustness to density,brightness,illumination and texture changes.