Multi-View Learning For Vehicle Re-Identification

Vehicle re-identification(re-id)aims to identify a target vehicle in different cameras with non-overlapping views,and it plays an important role in traffic surveillance scenarios when the car license plate recognition is unavailable or unreliable.Different camera views will have a great impact on the performance of vehicle re-id model.On the one hand,different views greatly affect the visual appearance of a vehicle,that is “intra-class variance”.On the other hand,different vehicles may exhibit fairly similar visual appearance when their images are captured from one unified single view,that is “inter-class similarity”To handle intra-class variance and inter-class similarity caused by view factors in vehicle re-id and improve the accuracy of vehicle re-id model,this thesis proposes a vehicle re-id method fused with latent views.The method incorporates a noval multi-view ranking loss(MRL)with an auxiliary classification loss to optimize CNN jointly.After that,our proposed method can extract view-invariant and discriminative vehicle features.Firstly,we introduce several groups to represent latent views which use to judge whether two images are captured from one unified single view.And then the latent views are incorporated into triplet constraint and vehicle re-id task is modeled as two sub tasks,including matching vehicles in a same view and across different views through metric learning.It makes the images of a vehicle more compact,and increases the distance between different vehicles meanwhile.Finally,the proposed method can easily distinguish different vehicles by consine similarity.In addition to the re-id accuracy,vehicle re-id task also has certain requirements for real-time capability,which means that vehicle re-id algorithm needs to quickly find the target vehicle from a large-scale dataset within a specified time.In order to solve this problem,this thesis proposes a vehicle re-id method based on a lightweight deep neural network.The proposed method uses Mobile Net V2 as the backbone network to extract vehicle features efficiently.Mobile Net V2 reduces the convolution operation and improves the efficiency of network through depthwise separable convolution.And inverted residuals block is proposed to fuse the image features and enhance the feature representation ability of CNN.In addition,the proposed method uses an adaptive weights sampling method to select triplets in the multi-view ranking loss to avoid the invalid training problem in randomly sampling,which can smooth the network traing process and accelerate the convergence speed of CNN.The proposed methods are evaluated on two public vehicle dataset.Extensive experimental results demonstrate that the proposed vehicle re-id method fused with latent views can identify more vehicle images from different views and achieve state-of-the-art performance.Otherwise,vehicle re-id method based on a lightweight deep neural network can greatly improve the efficiency of vehicle re-id method while keeping a high re-id accuracy.The delay time of this method is lower than vehicle re-id method fused with latent views.