Surveillance Video Analysis And Event Detection Based On Deep Learning

With the rapid development of deep learning,intelligent video surveillance systems have become current hot research issues,which involves computer vision tasks such as object detection,tracking,event detection and recognition.This thesis focuses on three research tasks:pedestrian detection,event detection,and crowd density estimation.For pedestrian detection,we improve the R-FCN detection framework from four aspects:model structure,proposal region generation,region feature extraction and hard sample mining,and implement a pedestrian detection model with both accuracy and speed advantages.The proposed method surpasses other existing methods on the SED-PD.v2 dataset and achieves comparable performance to state-of-the-art methods on the Caltech dataset.For event detection,aiming at the crowded surveillance scene in the SED-2017 evaluation,we propose an end-to-end individual event detection method,which directly uses the R-FCN model to locate the target event in multi-frame input.For the surveillance scenes with sparse event distribution and small related objects in the ActEV-201 8 evaluation,we propose a human-vehicle event detection method based on proposal segments,which first detects people and cars in videos and generates proposal segments,then locates the temporal boundaries through the action recognition model.These two proposed methods ranked in the forefront of the SED-2017 and ActEV-2018 evaluation,respectively.For crowd density estimation,we propose a scale aggregation network model based on two keypoints,multi-scale feature representation and high-resolution density map.The model consists of two parts:feature extraction module and density map estimation module.The feature extraction module is used to extract multi-scale features from the image,and the density map estimation module is applied to generate high-resolution density maps.In addition,in order to utilize the local correlation in density maps,we propose the local pattern consistency loss function to achieve multi-level supervised learning.This approach achieves superior performance to state-of-the-art methods on four major crowd counting datasets.