Research On Small Target Detection Method In Driving Environment Based On Improved YOLOv3 Network Model

Small target detection in the driving environment is a sub-problem of target detection,and it is one of the popular research directions in deep learning,computer vision and artificial intelligence.Compared with traditional image processing,the target detection network model based on deep learning has the advantages of high detection accuracy,strong real-time performance,good stability and support for simultaneous multi-target detection.With the continuous development of automobile intelligence,target detection is playing an increasingly important role in the field of autonomous driving.In the process of target detection,the detection of small targets has always been a difficult point in research.Only by solving the detection of small targets in the driving environment,can the target detection network model based on deep learning be widely used in automatic driving.This paper focuses on the detection of small targets in the driving environment based on the improved YOLOv3 network model.Based on the YOLOv3 network model,the shallow and deep networks of the backbone network of the network model are improved,and a small and medium driving environment based on the improved YOLOv3 network model is proposed.Target detection algorithm.First,analyze the target detection theory based on deep learning,including the working principle and detection process of the YOLOv3 network model,the overall network structure of the YOLOv3 network model,the normalization process,the calculation of the loss function,and the non-maximum suppression algorithm of the YOLOv3 network model,Lay the foundation for improving the YOLOv3 network model in subsequent chapters.Next,analyze the clustering prior frame process of the YOLOv3 network model,and propose using the improved K-means++ mean clustering algorithm to optimize the prior frame of the data set,thereby clustering the prior frame suitable for small targets in the driving environment.The experimental results show that the value of the objective function of the improved K-means++ clustering algorithm decreases faster than the original clustering algorithm,and the trend change is more stable,indicating that the improved K-means++ clustering algorithm can reduce the clustering deviation.And make a small target data set in the driving environment.The data set selects five categories of pictures of vehicles,pedestrians,left and right turning ground signs,straight ground signs and no traffic signs.Each category selects 500 samples as the training set,and each category Select 100 samples as the test set.The data set is used to test the detection performance of different network models.Then,aiming at the problem of poor feature extraction ability of YOLOv3 network model for small targets and small receptive fields,the improved M-YOLOv3 network model based on multi-scale fusion is proposed.The model adjusts the image input of the entire network,from the original 416×416 pixels to 512×512 pixels,increases the number of shallow convolutional networks in the backbone network of the YOLOv3 network model,and adds a layer of shallow output on the basis of the YOLOv3 network Layer,through the fusion of deep features,a total of 4 scales of prediction output are achieved.Experimental results show that the m AP value of the M-YOLOv3 network model is 5.43% higher than the original YOLOv3 network model,the detection accuracy of the network model is improved,and the improved network model still has strong real-time performance.Finally,on the basis of multi-scale fusion,an improved T-YOLOv3 network model based on DenseNet and multi-scale fusion is proposed to detect small targets in the driving environment.The T-YOLOv3 network model increases the number of shallow convolutional networks and output layers,and at the same time adds DenseNet dense convolutional networks to the deep layer of the Darknet-53 backbone network.The original transmission layer with lower resolution is used to weaken the deep network.The gradient disappears and enhances feature transfer and promotes feature reuse and fusion.Experimental results show that the m AP value of the T-YOLOv3 network model is 8.04% higher than the original YOLOv3 network model,2.61% higher than the M-YOLOv3 network model,and the average detection time is 0.025 s.Compared with the 0.023 s of the original YOLOv3,it still has a strong real-time performance.