Research On Small Object Detection In Road Environment

With the continuous development of deep learning,object detection has become a hot spot in academic research and engineering application in various fields.In different detection tasks,vehicles and pedestrians in the distance,flying objects in the sky and small animals can be regarded as small objects.These small objects have diverse categories,cross fields and variable scales,which have become the difficulties affecting the overall detection accuracy.Especially in the field of intelligent transportation,realtime and accurate detection of small objects can achieve early warning and avoid major traffic accidents.YOLOv4-Tiny detection model can meet the needs of real-time detection,but there is still room for improvement in detection accuracy.This thesis selects it as the basic model to further study the methods to enhance the features of small objects and improve the performance of small object detection from the perspectives of receptive field,feature fusion and attention mechanism.The main contents are as follows:(1)Dilated convolution fusion module based on receptive field: DCF.From the perspective of receptive field,the DCF module is designed: Firstly,the dilated convolution with dilated factor of 1 is used to increase the receptive field of the original feature map to obtain the enhanced feature map;Secondly,it is fused with the original feature map to obtain the fused feature map;Then,the dilated convolution with dilated factors of 2 and 3 is used to repeat the above two steps in turn to obtain a new feature map after increasing receptive field;Finally,the DCF module is inserted into the original model to get the improved model.Through experiments on public data sets and self collected data sets,the effectiveness of DCF module is verified,and the feature information of small objects is increased,and the object detection accuracy is improved.(2)Multi-scale feature fusion module based on DCF module: DCMF.Based on the enhanced feature map of DCF module,the DCMF module is designed from the perspective of multi-scale feature fusion: Firstly,the shallow feature map is down sampled and fused with the deep feature map,and a new deep feature map is obtained by aggregating the fused features with convolution kernel;Then,the new deep feature map is fused with the shallow feature map,and the convolution kernel is used to aggregate the fused features to obtain a new shallow feature map;Finally,DCMF module is used to replace the fusion structure in the original model to get the improved model.Through experiments on public data sets and self collected data sets,the effectiveness of DCMF module is verified,the feature extraction ability of small objects of DCF module is enhanced,and the object detection accuracy is improved.(3)Position sensitive module based on DCMF module: DCCAMF.From the perspective of spatial location information,in order to make network pay more attention to spatial features,Coordinate Attention is introduced to design DCCAMF module: Firstly,after three different dilated convolution of DCF module,the spatial features of the feature map behind the receptive field are decomposed according to the horizontal and vertical directions;Secondly,the location feature information in the horizontal and vertical directions is aggregated,and the Coordinate Attention is introduced to make the network learn the feature information adaptively;Then,combined with multi-scale feature fusion,the DCCAMF module is obtained;Finally,DCCAMF module is used to replace the fusion structure in the original model to get the improved model.Through experiments on public data sets and self collected data sets,the effectiveness of introducing Coordinate Attention is verified and the object detection accuracy is improved.This thesis improves YOLOv4-Tiny model: Firstly,based on the dilated convolution,DCF module is proposed to increase the receptive field of small objects;Secondly,using multi-scale feature fusion and proposing DCMF module to enhance the features extracted by DCF module and enrich the feature information;Then,DCCAMF module is proposed by introducing the Coordinate Attention to make the network pay more attention to the extracted features;Finally,the improved modules are verified by experiments and the experimental results are analyzed.Experiments show that the improved module proposed in this thesis is effective,the overall detection accuracy of the improved model based on the proposed module is improved to a certain extent.