Research On Robust Detection Algorithm Of Small-Scale Objects For Surveillance Video

In recent years,with the rapid development of deep learning,various tasks in computer vision have been widely studied,in which deep learning object detection is an important branche.At present,the object detection network and optimization strategy proposed by researchers have achieved a relatively robust effect on the detection of medium and large-scale objects.But the detection effect for small-scale objects is unsatisfactory.The detection rate of the existing object detection network is not high due to the characteristics of small absolute or relative size of small-scale objects,fuzzy features and easy to be confused with local areas in the background.In order to improve the detection rate of small-scale objects,in this thesis,small-scale UAV detection for surveillance video is taken as the background,and improvements are made from the following three aspects:1)A general attention model is proposed.Due to the objective factor that the features of small-scale object are ambiguity,this thesis proposes four improvements to enhance object features and improve the detection accuracy of small-scale objects on the basis of multi-head self-attention.First of all,the full connection layer in multihead self-attention is replaced by convolution module,so that the improved model is plug-and-play.Then,in view of the characteristics of fuzzy features,CBAM module is integrated into multi-head self-attention.By adjusting the input of self-attention,the characteristics of small-scale objects are initially enhanced.Next,for the characteristic that effective area of the object is small,the feature pyramid is integrated into CBAM,which improves the ability that the network searches for the effective features of small-scale objects;Finally,considering that shallow features contain more effective features of objects,the residual features are incorporated into CBAM to improve the reuse rate of shallow features.2)A YOLOv5 s network structure optimized for the characteristics of small-scale objects is proposed.Due to the small absolute or relative size of small-scale objects,the effective features that can be provided for network learning are very limited.A properly sized receptive field is selected,which can effectively assist the network to extract more object features,reduce the introduction of background redundant features,and improve the detection effect.In this thesis,the s-YOLOv5 s network structure is constructed by adjusting the YOLOv5 s network structure and improving the utilization of deep features.The s-YOLOv5 s structure can better learn the features of small-scale objects,resulting in higher detection accuracy.3)A lightweight research method for small-scale object detection network is proposed.At present,in order to improve the detection accuracy,the methods such as complex connection relationship,stacking network depth or increasing network width are usually used in most deep learning object detection algorithms.It generally results in a large number of parameters and calculations,which makes it difficult to deploy to edge computing devices.Thus,in this thesis,a lightweight detection method is proposed.The lightweight strategy is realized by the idea of first lightening the model and then improving the accuracy.First of all,considering that the parameters of convolutional neural network are basically concentrated on convolutional kernels,in this thesis,the number of convolution kernels is greatly reduced on the basis of sYOLOv5 s,which realizes the lightweight of the model.Then,considering that shallow features contain more effective object features,three non-parametric feature reconstruction models are designed to improve the utilization rate of shallow features and the fusion rate of shallow features and deep features,so as to build more robust semantic features and improve the accuracy of small-scale object detection.