Research On Retained Object Recognition And Abnormal Behavior Detection Based On Scene Understanding And Deep Learning

Retained object detection and recognition is one of the most important research topics in the field of computer vision.In public places and crowded areas,undetected dropped object may threaten pedestrians’ lives and property.Therefore,intelligent monitoring and automatic recognition of video are required.In real life,when stranded objects appear in a scene,they can be detected and identified quickly and can be given effective early warning,which plays a great role in urban management and security.Complex scene problems,such as background complexity,light change,low quality image and occlusion,make the detection and identification of retained objects very challenging.Meanwhile,the detection of retained objects lacks the correlation analysis of its owner,and there is no good solution for the classification of low-quality images and videos.Therefore,this dissertation proposes multiple retained object detection and recognition methods from traditional methods and deep learning methods.Main tasks are as follows:(1)Aiming at the correlation between the object and the pedestrian,a target tracking algorithm is proposed based on the connected region information.The relevant retained object algorithm only focuses on the object itself,lacks the analysis of the pedestrian of its correlation,which cannot obtain the owner of the retention,and it is also lacking the category analysis of the retained object.Therefore,the algorithm uses the improved correlation filtering algorithm for pedestrian tracking,multi-directional background model for pedestrian and static target detection,and uses the connected area at the time of pedestrian and object separation to detect the retained object.At the same time,the combination of principal component analysis and Euclidean distance was used to identify the retained object.Proximity algorithm combining geometric affine invariant matrix and Euclidean distance matches the retentate with the sample images in the item library to obtain the recognition result.The recognition results may be applied to further analyze human activities and intentions such as determining whether the retained object are intentional hazardous or unconsciously lost articles according to the types of retained object.Good performance has been achieved on the public datasets of VISOR and CAVIAR.(2)Aiming at the static targets in complex scenes,which have different meanings in different scenes and lack of multi-scale judgment,this paper proposes a stranded object detection algorithm based on scene understanding.Because the scenes of retained object detection are mostly the scenes with high crowd and complex background,and static targets have different meanings in different scenes and at different times,they are not necessarily the targets of retention and harm.So a multi-feature fusion algorithm is used to detect static objects at first.Static objects are then re-detected by a Deep Learning method to determine the location and class of the dropped object.Finally,an efficient semantic information method relying on scene parsing and classification information is applied to determine dropped objects.At the same time,the abnormal behavior of the owner can be analyzed by semantic judgment.The algorithm has achieved good results in PETS2007,PETS2006,CDNET2014 and ABODA datasets,and it has high detection accuracy in different complex scenes and strong generalization ability,(3)Aiming at the classification of objects in low-quality videos and images and the classification of objects with incomplete target information,a fine-grained classification algorithm based on image superresolution reconstruction is proposed.For different monitoring equipment,the quality of video shot and image is different,especially in large public places,it is difficult for the shooting equipment to include all the content in the distant view and scene into the video and image,which will lead to incomplete information collection and affect the identification of the target.First,image quality analysis is performed,and multiple evaluation indicators are used to analyze the picture.Then perform super-resolution image reconstruction to improve image quality.Finally,a compact bilinear pooling fine-grained item classification algorithm is used to train and detect the dropped object data we have established.Experiments show that image quality has a huge impact on fine-grained item classification algorithms.Improving image quality through image super-resolution algorithms can improve the success rate of fine-grained item classification.The PETS2006,2007 and ABODA data sets were used to compare the classification of items with different quality images,and the classification accuracy was better with high quality images.