Research On Specific Target Vehicle Detection And Matching Algorithm Based On Neural Network

In recent years,with the increasing number of automobiles in our country year by year,the intelligent transportation system that makes full use of high-tech to solve traffic problems has received more and more attention.As one of the important functions of the intelligent transportation system,the accurate recognition of road vehicle targets can be regarded as the target matching of specific vehicles in road video images.Accurate,stable,real-time,and fast detection and matching of road target vehicles is crucial for intelligent transportation systems.The vehicle detection and specific vehicle target matching method proposed in this paper is mainly dedicated to the effective detection of vehicle targets on the road in traffic scenarios and the effective matching of specific target vehicles that are expected to be identified,such as illegal parking.Attribute information and precise positioning provide effective data support and technical support for the intelligent transportation system.The technical research in this paper mainly includes the YOLOv5 network vehicle target detection method based on SE（Squeeze and Excitation）attention mechanism and the specific vehicle matching algorithm based on sketch conversion,template matching and Siamese network fusion.Among them,the YOLOv5 network vehicle target detection based on the SE attention mechanism first introduces the SE channel attention mechanism in the YOLOv5 model.After training the model,the vehicle targets are detected and obtained.Then the vehicle feature information is enhanced by converting the vehicle image to its sketch,which is beneficial to subsequent matching.Furthermore the template matching method is applied to roughly match the specific target vehicle in the road video image containing many vehicles.The Siamese network is used to fine match the matched vehicle,and the final accurate specific vehicle matching results can be obtained.The method flow is as follows.（1）First,for the initial YOLOv5 model,the SE attention mechanism is introduced into the residual network module.After inputting the training set vehicle image into the residual module,the first step is to obtain two compressed feature maps through the global pooling extrusion operation and send the obtained feature maps into two layers of neural networks with shared weights.After activated by the activation function Re LU,the features output by the two layers of neural networks are added element by element,and then activated by the sigmoid function to obtain the final channel attention weight feature map.The SE attention mechanism strengthens the feature learning during training of the YOLOv5 model by adding a global average pooling and two fully connected layers and feature weighting.It obtains a good m AP improvement on the validation set.After the training is completed,the trained YOLOv5 model is used to detect the monitored road image after frame extraction,and the accurate vehicle target result after detection is obtained,and the subsequent specific target vehicle matching is performed.（2）After obtaining the detected vehicle target results,the specific target vehicle of violation to be matched is screened and determined.After the matching target is determined,the image of the specific target vehicle to be matched and the original road to be matched are processed through the sketch conversion method based on the Canny operator firstly to enhance the vehicle feature information.After the sketch images are obtained,the template matching method is applied to roughly match the specific target vehicle in the original road image which contains many vehicles to be matched.The trained Siamese network performs fine matching on the matching vehicle results and obtains the final and accurate specific vehicle matching results.This matching method has a high matching accuracy rate on the validation set,and effectively achieves the precise matching of specific illegal target vehicles in complex traffic environments.This paper uses the test set data of road surveillance video acquired by Changchun Boli Electronic Technology Company for experimental verification.First the trained YOLOv5 model with SE attention mechanism is used to detect the target vehicle in the test set vehicle images and compare with network models such as YOLOv5 and YOLOv3.The experimental results verify the higher vehicle detection accuracy of the detection model in this paper.Then,the vehicle target of the test set is extracted by using the trained model,and the specific target matching vehicle is determined.Next the target vehicle and the image to be matched are matched by the image matching method based on sketch transformation and the optimized Siamese network proposed in this paper.By comparing with the methods of template matching and fast similar image template matching based on robust features,the matching advantages of the proposed method are verified,and the matching experimental results are given,which provides an important reference for the follow-up research of vehicle recognition and tracking.