Motion Foreground Detection Based On Time Series Feature Analysis

In recent years,video surveillance has played an increasingly important role in various fields such as road traffic,public security,industrial production,etc.How to efficiently obtain valuable sports information automatically from massive surveillance video data has attracted extensive attention from academia and industry.Researchers have proposed a large number of motion foreground detection algorithms,but these traditional methods will encounter many challenges when applied to actual complex and changeable scenes,such as the impact of lighting changes,camera shake,and severe weather conditions.The motion foreground detection algorithm based on deep learning improves the robustness of the detection to a certain extent.However,the current deep learning models usually only use spatial features to achieve motion feature extraction,and to varying degrees ignore the timing features that better reflect the essence of motion modeling.Therefore,how to design an algorithm that makes full use of time-series features for motion foreground detection is of great research significance for coping with various challenges in real scenes and improving the promotion ability of the model.This paper starts from improving the robustness of motion foreground detection to various scenes,fully exploits the time-series features of video,and uses deep neural networks to analyze and learn the motion information of objects to realize the detection of motion foreground.The main work of the paper is:(1)In view of the existing deep learning models usually only use spatial features to realize the current situation of motion foreground detection,this paper builds a 3D CNN-Conv LSTM motion foreground detection algorithm based on time series features.The model of the algorithm mainly includes two parts.First,the model uses a three-dimensional convolutional neural network(3D CNN)that is widely used in the field of motion recognition to extract time-series features.Using the characteristics of 3D CNN to perform convolution operations in the time dimension,to learn the time-series features in a short time.At the same time,in order to be able to train the model better,a residual network structure is used in the 3D CNN.Then,in order to learn the features of a longer time range,the Conv LSTM network that can consider the spatial correlation is selected and used.Since a large amount of time and space information is added,the Focal Loss function is used to solve the imbalance problem in which the number of background pixels in the data is much larger than the number of foreground pixels.Experiments show that it is feasible to use temporal features to detect motion foreground,and the algorithm has a lot of room for improvement,so consider using more abundant timing features to improve the performance of the algorithm.(2)In order to improve the motion foreground detection of difficult scenes with the help of richer time series features,this paper combines the 3D convolutional neural network in Chapter 3 to construct a 3D Atrous CNN algorithm to realize the detection of motion foreground.Simultaneously use three-dimensional hole convolution in the time dimension and space dimension to capture the time series features of a longer time span and learn the motion information in a larger field of view.In addition,using multiple scales in time,feature maps with different time dimensions are generated by using different time atrous rates and video frame input step sizes.In this way,the model has a stronger ability to express target motion,which is more conducive to the robustness of motion foreground detection.Experiments show that after adding multi-resolution timing features to the 3D CNN,compared with the algorithm in Chapter 3,the performance of the algorithm is improved by nearly 20%,and the accuracy rate of 96% is reached.Fully explain the effectiveness of the improved algorithm model.