| With the further development of deep learning research,based on the deep learning target detection technology can effectively avoid the common problems existing in daily life,at the same time some target detection algorithm is not applied in actual situation,the reason is that the detection speed is too slow,not very good real-time detection ability,cost a lot of computer resources.So this paper mainly for running faster One Stage target detection algorithm-YOLOv3,YOLOv3 is a classic model of One Stage,the algorithm based on the end-to-end target detection idea,increase the anchor mechanism and using multi-scale fusion prediction to realize the positioning and classification of target detection task,belongs to a kind of better target detection algorithm.The main research contents of this paper include:Firstly,the selected urban road data set(SODA 10M)is amplified to improve the detection accuracy,several methods used in deep learning are introduced and corresponding examples;two relatively complex data enhancement methods for mixing two images are described;and the enhanced data set of Simple Copy-pace(SCP)method is the data set used in this paper to improve the accuracy of road detection.Secondly,for the partial reinforcement of the backbone network feature extraction,using a network combining lightweight Mobilenet-v2 and YOLOv3 making the entire network run faster,with the effect of optimizing network structure and improving feature extraction ability;meanwhile,replace a loss function that makes the training more balanced,called the Balanced-l1-loss,of the training for faster convergence by adjusting the parameters,the final test is better;finally,using the AF K-MC2 algorithm,it is a new clustering algorithm used to generate the anchor prediction boxes,in contrast to the K-means clustering algorithm in the original network,it does not depend on the selection of the initial box and does not rely on any assumptions,thus is called the non-hypothetical markov chain clustering algorithm,each step of it is merely related to its previous step,it is a very good algorithm for generating predictive boxes and achieve a better predictive effect.Last,the network to reach 58.8%mAP on the representative urban road augmented data set(SODA 10M),compared to the improved detection speed,the detection accuracy was also improved by about 12.4%,and comparing the detection accuracy and detection speed with some excellent network models of YOLO series,finally found that our improved detection results of the network model are better,the application of road detection has better real-time detection effect,can try to apply to road condition detection in actual traffic systems,it has a relatively fast detection speed,applicable in the intelligent transportation system. |
