Research On Vehicle Re-identification Based Convolutional Neural Network

In recent years,smart transportation has developed rapidly,and the development of autonomous driving is getting faster.The demands of autonomous driving and vehicle-road coordination have accelerated the new traffic pattern.Meanwhile,road safety has attracted people’s attention.Vehicle re-recognition can realize the positioning and tracking of the vehicle’s cross-camera,which is of great significance to road safety and intelligent transportation.The essence of the vehicle re-recognition problem is target re-recognition,but it is different from the classic pedestrian re-recognition.As many vehicles of the same model have similar visual appearances,re-recognition of vehicles is more difficult.Especially when the license plate cannot be obtained effectively,it is harder for people to distinguish the same type of the cars.Therefore,the vehicle re-recognition algorithm needs to find the differences between and within the class at the same time.While the license plate cannot be obtained effectively,the method proposed in this paper can be used to extract the characteristics of the vehicle model,vehicle color,window inspection mark,body painting and even scratches to solve the problem of vehicle re-recognition.It is proposes that a vehicle re-recognition algorithm based on an improved deep relative distance learning model.Firstly,for the feature extraction network in the original deep relative distance learning framework is simple and the vehicle features cannot be better extracted,a Rep Net network is proposed to replace the VGG CNN M2048 network in the original network architecture.After the improvement,the coarse-grained learning channel responsible for label attribute classification first extract the color and vehicle model of the vehicle,and then feedback the extracted features to the subsequent fine-grained similarity learning through the suppression layer to eliminate embedding into the fine-grained Coarse-grained features.It not only enables the model to focus more on the extraction of complex features during fine-grained learning,but also improves the recognition accuracy.There is an obvious imbalance of positive and negative samples in the vehicle re-recognition database which lead to lower training efficiency and may even degrade the performance of our trained model.In this paper,focal loss function is proposed for this situation,which can reduce the weight of simple samples in training.After experimental verification,in the classification of vehicle label attributes,the recognition rate of the model designed is 98.18%,which is about 14.7% higher than the original DRDL model.The recognition rate of vehicle color reaches 96.28%.In the vehicle re-recognition task,the MAP value of the model adopted reaches 0.709,which is about0.16 higher than the original model.The results prove that the effectiveness of the method proposed in this paper.