Research On Vehicle Re-Identification Based On Deep Learning

In recent years,with the development of "smart cities" and "smart transportation",thousands of surveillance cameras have been deployed in every corner of the city,massive video data has explosively increased,the demand for smart video surveillance has grown rapidly,and smart video surveillance systems become an important part of the construction in smart transportation and safe city.Vehicle re-identification is one of the core technologies in the intelligent video surveillance system,which can achieve cross-camera tracking,driving trajectory prediction,and positioning of target vehicles.It has a wide range of practical applications and important research value.However,vehicle re-identification is a very challenging task.In urban environments,vehicle re-identification often suffers from various challenges,such as viewpoint variations,illumination changes,object occlusions,complicated surveillance scenarios,especially the appearances of vehicles of the same vehicle model and color are very similar.With the rapid development of deep learning technology,vehicle re-identification has made great progress in recent years.Therefore,by concentrating on vehicle characteristics and the difficulties of vehicle re-identification,this thesis is to study deep learning-based vehicle re-identification algorithms.The main research contents are as follows.On the one hand,a multi-label learning with multi-label smoothing regularization algorithm for vehicle re-identification is proposed.Based on the observation that essential vehicle attributes,like vehicle‘s colors and types,could be used as important traits to recognize the vehicle,an effective multi-label learning(MLL)method is proposed,which can conduct feature learning according to three labels(i.e.,vehicle’s ID,type,and color)simultaneously.With three labels,a multi-label smoothing regularization(MLSR)is further proposed,which can allocate a uniform label distribution to the multi-labeled training images to regularize MLL model and improve vehicle re-identification performance.Extensive experiments have been conducted on two public and challenging vehicle re-identification datasets,demonstrating that the proposed method consistently outperforms multiple state-of-the-art vehicle re-identification methods.On the other hand,a deep quadruplet appearance learning method for vehicle re-identi?cation is proposed.Since vehicles with the same model and color have very similar vehicle appearances,which poses a huge challenge for vehicle re-identification.For that,the proposed method designs the concept of quadruplet and forms the quadruplets as the input,where each quadruplet is composed of the anchor,positive,negative,and the specially considered high-similar(i.e.,the same model and color but different IDs with respect to the anchor)vehicle samples.Then,the quadruplet network with the incorporation of the proposed quadruplet loss and softmax loss is developed to learn a more discriminative feature for vehicle re-identi?cation,the mapping of the quadruplet to the Euclidean space through the quadruplet network model should satisfy the following distance relationship: the distance from the positive sample to the anchor less than the distance from the high similarity to the anchor,and less than the distance from the negative sample to the anchor.Extensive experiments demonstrate that the proposed method can further improve the performance of vehicle re-identification and consistently outperforms multiple state-of-the-art vehicle re-identification methods.In summary,to some extent,the proposed methods in this thesis broaden the thinking for vehicle re-identification and provide the technical support for the practical application of the vehicle intelligent surveillance systems.