The Research Of Object Recognition In Surveillance Scene

In recent years,camera equipments and other video devices have been widely used in reality,and the amount of video data captured has also increased significantly.Correspondingly,there is an increasing demand for the processing and analysis system of video data.Such as the abstraction system of surveillance video based on the object,which regards the pedestrian and vehicle in the video as the basic unit of the system,extracting the unique information and events of each object.The object recognition technology in the surveillance scene is the basis of the subsequent information extraction and analysis of the video analysis system,such as the enhancement of tracking robustness,the extraction of detail information,the exception analysis and other tasks,which depend on the result and performance of the object recognition.For the surveillance video abstraction system,with the framework of the robust monitoring video moving object detection and tracking algorithm based on background extraction,this paper mainly completes the three object recognition tasks which need to be completed or improved in the surveillance scene,object image classification,car type recognition and person re-identification:1.For the problem of image classification in the surveillance scene,this paper uses the CNN model which is trained on the large-scale classification task data set as the feature extractor to produce feature of the image.Then we propose an extended sparse model which learns a dictionary from the CNN features by incorporating the reconstruction residual term and the co-efficients adjustment term.Minimizing the reconstruction residual term guarantees that the class-specific sub-dictionary has good representation power for the samples from the corresponding class and minimizing the coefficients adjustment term encourages samples from different classes to be reconstructed by different class-specific sub-dictionaries.With this learned dictionary,the classification performance is improved.Finally,a metric involving these discriminative information is introduced for image classification in the surveillance scene.Experiments on Caltech-101 and VOC 2012 datasets validate the performance of the image classification framework.2.For the problem of car type recognition in the surveillance scene,considering the practicability of the target surveillance video abstraction system,this paper sorts out the classification datasets of vehicle for surveillance scene from the Comp Car dataset,which containing a large number of general models of Chinese mainland.In this paper,we use the Alex Net,Google Net and VGGNet of CNN model which is suitable for transfer classification task as the basic model.The fine-tuning experimental results and the characteristics of the car data show that,the frontside image and local image of cars contain recognition is classification information for the car recognition task.In this paper,a multi-task neural network model for car type task is proposed,a dual output neural network model is established,which is the car type recognition and the front,front-side and the local view recognition task.Through the training of the model,the accuracy rate of classification task is improved.The comparison experiments verifies the validity of the car type recognition algorithm.3.For the problem of person re-identification,this paper treats the problem as a retrieval problem and tries to find an effective measurement method by metric learning,which makes it possible to calculate the distance from the pedestrian image in the sample set by using the measure method when calculating the pedestrian images belonging to the same object can be ranked in the minimum similarity distance or the previous minimum similarity distance results.The projection matrix utilizes the method of mapping samples with different spatial distribution to a coupled public space with a more easily metric distance,which is inspired by this method,and constructs a generalized projection metric matrix based on the autocorrelation and correlation metrics of samples of spatial distributions.The similarity loss function is constructed by using the training sample selection method of ternary group,which eliminates the imbalance between matching and mismatched samples.Finally,we use LOMO feature to describe the pedestrian sample,which incorporates the color and texture information of the pedestrian image.Experiments on the i-LIDS and VIPe R datasets verifies the performance the algorithm.