Research On Vehicle Detection And Counting Based On Feature Anti-Interference And Scale Context

Vehicle detection and counting based on remote sensing images of UAVs and satellites are increasingly widely applied.With the development of deep learning and the update of computer hardware equipment,vehicle detection and counting methods based on deep learning have become a key research direction because of their advantages of high precision and good robustness.The vehicle detection method based on regression density map can classify each pixel and increase the utilization of image information.However,the existing regression density map methods,such as FCRN,SDBN,PRFN,etc.,have some common shortcomings:(1)Lack of feature anti-interference ability,easy to be affected by complex background features.(2)The acquisition of context information of small-scale vehicles is not sufficient.(3)The vehicle target information loss in feature extraction is not fully considered.To solve the above problems,a vehicle detection network FICLAR-Net is proposed in this paper.The innovations are as follows:(1)A feature interference module is designed to generate interference features into the detection network,and the network is trained to resist the influence of useless features,so as to enhance its feature anti-interference ability.(2)An adaptive residual attention module is constructed,features are extracted by deforming convolution adaptive to improve the network’s attention to occluded targets,and the pixel attention mechanism is combined with the channel attention mechanism to obtain context information.(3)A cross level fusion module is designed to extract multi-layer features to reduce the loss of feature information of vehicle targets.Ablation and comparison experiments are conducted on UCAS-AOD,CARPK and OVDS for the proposed detection network FICLAR-Net.Compared with the mainstream methods,the F1-score of the proposed network increased by 0.47%,0.81% and 0.37%,respectively.Thus,the validity of the proposed detection method is verified.The vehicle counting method based on density estimation has high counting accuracy and can display the target distribution information through the output density map.However,the current density estimating-based counting methods still have some shortcomings:(1)The existing methods are dominated by the fully supervised learning method.Although the existing weak supervised methods reduce the use of label information,they still need to manually generate density map labels.(2)The density map output by the current method can only display the general distribution information of vehicles,and cannot locate vehicle targets.To solve the above problems,a weakly supervised vehicle counting network SCRC-Net is proposed in this paper.The innovations are as follows:(1)The network only needs the number information of vehicle targets during training,thus reducing the tagging cost.(2)The scale context module is designed in this paper.Multi-scale features are screened by designing a dual feature screening structure,and context modeling is completed by combining the convolution structure with Transformer.In addition,the region constraint network is constructed,and the regional constraint loss is designed to guide the detection region to concentrate on the target region,and the more accurate density map are returned to complete the vehicle center positioning.In this paper,the ablation experiment and comparison experiment of the proposed detection network SCRC-Net were carried out on three public data sets of PUCPR+,CARPK and Drone Vehicle.Compared with the mainstream methods,the MSE of the proposed network was increased by 0.5,0.4 and 0.7respectively while the use of label information was reduced.Thus,the validity of the proposed counting method is verified.Finally,the work content of remote sensing image vehicle detection and weakly supervised vehicle counting is summarized,and the future research direction of remote sensing image vehicle detection and vehicle counting is prospected.