| Pattern recognition based on convolutional neural networks has been widely used,and it is a typical application example for vehicle target recognition in video.In this paper,the YOLO model is used to identify the vehicle target.In the process of identification,when multiple candidate frames overlap,the confidence of the overlapping candidate frames is compared.When a certain threshold is reached,non-maximum suppression is triggered to remove multiple candidates belonging to the same category target.A box of non-maximum confidence in the box.However,the traditional non-maximum suppression has shortcomings,and it is impossible to suppress the target with large difference in area.Therefore,for the union of the intersection ratio for non-maximum suppression,the minimum area of two overlapping target regions is used instead.,to achieve the effect of triggering non-maximum suppression.In this paper,the verification phase of YOLO model is added to the improved non-maximum suppression algorithm,and vehicle identification is taken as the theoretical verification background of this paper.The superiority of the improved model proposed in this paper is verified.After the vehicle identification is completed,the number of vehicles can theoretically be obtained.Then the neural network is used to predict the traffic flow.Considering that the learning rate of the traditional neural network model is based on artificial empirical setting,the value of the learning rate is set too large,which tends to cause the model to oscillate during the training process;and the learning rate is set too small,which tends to cause the model gradient to slow down slowly.The convergence speed is slow,so it takes a lot of time and energy to make the model converge.Obviously this is a disadvantage caused by artificially setting the learning rate.Therefore,this paper proposes an adaptive learning rate optimization method based on gradient change.Gradient is used as the basis for judging the new learning rate.Each time the error propagates back,a new learning rate is generated.The change of the learning rate is completely proportional to the absolute value of the gradient change.When the gradient changes greatly,the learning rate is larger.Value,speed up the model training;when the model is close to the ideal optimal,it is the valley of the gradient of the model training,the gradient of the model is relatively flat,the gradient change is small,the learning rate is small,the training speed is reduced,and more Helps the model converge to the point where the error is at a minimum.In this paper,LSTM,GRU and SAEs models are used to verify the model with adaptive learning rate.The traffic flow prediction is used as the application background and compared with the traditional unoptimized model.Experiments show that the improved YOLO model can accurately determine the category of vehicle targets.A single target will only get a candidate box.Improving the traditional model will produce the disadvantages of multiple candidate frames.In this paper,the LSTM,GRU,SAEs model is used to fit the traffic flow,and the adaptive learning rate model is adopted.Compared with the unoptimized traditional model,it has smaller fitting error and higher correlation coefficient. |
