| Automatic driving technology can not only reduce the burden of drivers,but also reduce the occurrence of traffic accidents effectively,so it is of great significance.In recent years,with the breakthrough of convolutional neural network technology,new ideas have been provided to solve the key problems in autonomous driving.Through target detection method based on CNN,so cars can identify car,bicycles,traffic obstacles and a series of dynamic or static traffic targets accurately and timely and provides technical support for automatic driving.Therefore,design an effective target detection algorithm is very important for cars to obtain the information of the surrounding environment quickly and accurately.This paper is based on the YOLO-V3 real-time target detection algorithm and studies it in order to improve the recognition effect in the autonomous driving scene.The main work includes:1.This paper collected the images of four kinds of dynamic targets,such as ebike,car,person and bicycle,which are common on the road.In the collection process,a multi-angle and multi-distance collection strategy was adopted,combined with data set expansion,the paper complete the labeling of all images.2.Under the original size of anchor frame,the YOLO-V3 model is insensitive to the target size and has a low detection accuracy problem.This paper proposes a method of clustering target borders by K-means method.Firstly,the distance formula is improved in the process of clustering,then studied the clustering algorithm of K value relations with ious and K value is determined by mountain climbing method.So the experiment obtained nine new anchors.By replacing the original anchors,the experiment completed preliminary improvement.Finally,according to the training process designed in the paper,the paper completed two training tasks.The experimental results show that the F1 of the new anchors model is improved,and the accuracy of target detection is improved by 3.4%.3.Aiming at the problem that the model detection speed is slow and not suitable for road real-time detection,this paper proposes a method to merge the parameters of the BN layer into the convolutional layer.Firstly,the paper analyzed the workflow of the BN layer in the network and pointed out the problem of the BN layer.Then,the paper deduced the feasibility of merging parameters of BN layer to convolution layer.Finally,according to the test process designed by the paper,two comparative experiments are completed.The experimental results show that this method can improve the detection speed of YOLO-V3 model while keeping the detection accuracy unchanged,the increase rate was 14%.The FPS of 26.5 was obtained in the actual video test environment,which met the real-time detection requirements.4.In order to study the improvement effect of transfer learning based on this paper on the detection task in this paper,the experiment compared the training results of YOLO-V3 model under transfer training and non-transfer training.Experimental comparison results show that the method based on transfer training in this paper not only improves the training efficiency,but also increased the accuracy of target detection by 11.6%.For the detection task in this paper,the use of transfer training is effective.Figure[51]Table[19]Reference[79]... |
