Research On Key Problems In Vehicle Target Identification And Classification Based On Surveillance In Large City

Nowadays,video surveillance plays a more and more important role in security control and detection.How to make better use of monitoring images to manage public affairs and protect public safety has become a hot issue of public safety departments in various countries.Differ from the full swing development of human face recognition technology and large public data sets continuously built and enriched like LFW,vehicle——a target which has a same important position in public safety control as person target,is lagging behind researchs on human face identification because of the excessive dependence on plate license recognition and the lack of vehicle public datasets.However,in the real scenes of investigation,how to recognize what you want and how to classify all attributes of the vehicles from the massive video resources to identify relevant objectives,how to reduce the scope of investigation quickly in the image through the description of the eyewitness are the technologies extremely needed but lack of reaseach.In order to solve these difficult problems,this paper starts with the establishment of professional vehicle sample database,and conducts in-depth research mainly from two key directions:fine-grained vehicle type recognition and vehicle re-identification.The specific work and innovation points are as follows:1)In this paper,two sets of vehicle sample collecting schemes with different characteristics are designed.The design of collectings of balanced positive and negative samples and real unbalanced samples separately are very important for the normalization of sample data and the correctness of sample attribute determination.2)We construct a professional dataset with a total of 2,800 thousand of the vehicle pictures taken for 280 thousand of different vehicles.They can be divided into 1797 vehicle types in total and covered 90%vehicle types registered in Shanghai.These images are carefully and standardly classified in a multi-angle and multi-property manner.It surpasses the existing databases in terms of the number,type,standardization and classification grained.3)Firstly,we proposed an improved space-based vehicle type recognition algorithm.In order to solve the practical problems of most previous researches which are only for front vehicle recognition or low rate in accuracy,a new vehicle type recognition method from the side view is proposed.The method combins the advantages of multi algorithms,further enhance the discriminability of the data and the robustness of pose change,and achieves a high accuracy on a small dataset and lays the foundation for the full angle recognition of the later algorithm and the expansion of the medium or large scaled dataset.4)We propose an improved PCAnet network learning algorithm,resulting in greatly reduced pre-processing workloads at the early stage.Compared with other neural network algorithms,the proposed method adopts fewer network layers and substantially reduces the number of training sam-ples needed for network convergence;op-timal trade-off has been achieved for fewer computational loads,greater calculating efficiency,and higher recognition rate compared with popular CNNs.Results based on the new dataset constructed in this paper and other open datasets demonstrate the superiority of the proposed algorithm,and also indicate that the proposed algorithm is particularly effective for the medium and large datasets and the dataset of extremely similar vehicle types.Hence,the proposed al-gorithm is well suited for practical applications.5)We proposed a vehicle re-identification(Re-id)solution based on model of bag of visual words(BoVW).In order to solve the application requirements of the cross-camera searchs among massive images for a specific vehicle target,we solve problems of pose changes.Based on the research of original BoVW,we improve the key steps of the model,and achieve result which is adaptable to different scales,different positions and angles,and has acceptable lable work and time-comsuming.6)We proposed an angle adaptive vehicle re-identification algorithm based on metric learning of fusion features.Based on the algorithm of improved BoVW modle and taking the characteristics of vehicle datasets into account,an angle adaptive vehicle re-identification algorithm based on feature fusion and metric learning is proposed.Studies are performed on how to fuse multi-property features,introduce the location of features,address data imbalance,and choose between computational complexity and high identification performance.An improved pairwise metric learning algorithm characterized by high identification rate and low computational complexity,together with the CNN network-based triplet deep learning algorithm with high identification rate,are adopted.The solution selections between accuracy and calculation cost are put forward.7)We put forward a method of vehicle re-identification and vehicle attribute classification based on dropout CNN network.In allusion to the problem of the time consumption in pre-labling and preprocessing process,the unadaptable of environment changes and unbalanced samples,the unsupervised deep learning using multi-layers CNN algorithm for cross-cameras large scaled dataset is realized.At the same time,the added dropout parameters faster the convergence speed,significantly reduced training time,and get significantly improve in accuracy compared with the original network.The experimental results in two application directions of re-identification and attribute classification like vehicle model prove that the algorithm is a general algorithm for both image recognition and classification.The researchs of this paper expand a new pattern of vehicle identification and attributes classification,and provide real and available tools in investigation through surveilance,which greatly promotes the further development of video and image analysis technology in public security field.