Research On Video Anomaly Detection Methods Based On Genetic Programming

Video anomaly detection is becoming more and more important with the improving popularization of camera surveillance.One of the main methods is to extract the features with semantic information from the original video data and analyze the related behavior.Two commonly used video feature methods are presently used:manual design feature descriptor(low-level feature)and deep learning feature.For two different ways,many related research work has been proposed and promoted.At present,due to the inadequacy of motion description ability,artificial design features often fail to describe complex motion in crowded video scenes,which limits the performance of such methods.The method based on the characteristics of deep learning has been widely used in recent years and has achieved reasonable results.However,the deep learning network relies on complex network structure and has poor interpretability.At same time,because the network structure needs to be designed and adjusted manually,the effect of detection can not be guaranteed.As one of the evolutionary computing methods,genetic programming(GP)has been widely used in large-scale optimization and image analysis in recent years,and has unique advantages such as parallelism and strong interpretability.Nevertheless,the methods and research works of video anomaly detection based on genetic programming make use of the advantages of genetic programming to solve the problems in video anomaly detection,which is a relatively new method.Based on genetic programming method,this paper focus on the problems of the description ability of common artificial design features in video anomaly detection,the structure and parameters of deep learning network which need to be adjusted manually.Some related algorithms and improvement schemes are proposed.The main contribution are as follows:Propose an anomaly detection method of genetic programming based on weighted multidimensional optical flow:In video anomaly detection,continuous motion information is important.Aiming at the improvement of feature description ability of setting related parameters manually,a new feature weighted multi-dimensional optical flow based on histogram of optical flow is proposed.Furthermore,feature selection and optimization of related parameters are carried out by using the randomness of genetic programming,and a genetic programming method based on weighted multidimensional optical flow is proposed.Experiments on UMN datasets show that this feature descriptor has advantages over traditional manual feature descriptors in detection accuracy,and the detection accuracy of this method has defeated other anomaly detection methods.Propose an anomaly detection method of genetic programming based on evolutionary deep learning:Aiming at improvement of poor interpretability of deep learning method,an evolutionary deep learning method based on genetic programming is proposed.Firstly,it is experimented that spatial structure information plays an important role in anomaly detection.In this paper,the traditional feature descriptors are further improved,and a weighted spatial limited histogram of optical flow is proposed to further improve the descriptive ability of traditional artificial design features.This paper combines the idea of deep learning and uses genetic programming to overcome the limitations of manual construction and low interpretation of neural network,and proposes an evolutionary deep learning method based on genetic programming.The effect is the same as that of the mainstream deep learning methods.Meanwhile,the working areas of convolution kernels are defined and the important types of motion features are pointed out,which are consistent with the actual anomalies,and the process of automatic feature extraction and evolutionary learning is explained.Experiments on UCSD and Avenue datasets show that the detection accuracy of this method reaches the same level as that of the mainstream deep learning methods.At the same time.the working areas of convolution kernels are defined,and the important types of motion features are pointed out,which are consistent with the actual anomalies.It also explains the process of automatic feature extraction and evolutionary learning.Propose an anomaly detection method based on CNN designed by cartesian genetic programming:The structure and parameters of deep learning network need to be adjusted manually is proposed in the pervious contributions,then method of automatically designing CNN based on Cartesian genetic programming and applyed it to video anomaly detection is proposed.Deep learning has the advantages in extracting spatial structure information,and does not need prior knowledge.But designing the structure and parameters of the neural network requires a lot of experiments and trialand-error.In this paper,Cartesian genetic programming is used to design the structure of CNN,and a better CNN network is designed automatically.In the process of the anomaly detection,optical flow information is also used as a supplementary feature.The proposed method solves the problem that network structure needs manual design and achieves good results.Experiments on Subway datasets show that the detection accuracy of this method is higher than that of the mainstream deep learning methods.