Violence Detection Method Based On Deep Learning

With the rapidly development of science and technology,our lives become more convenient and safe.Meanwhile,intelligence and informatization have become an inevitable trend of social development.In recent years,utilizing monitoring equipment to detect,analyze and process human's behaviors is increasingly favored by researchers.In this type of researches,detecting and early warning of violence play extremely important roles in maintaining social order and protecting personal life and property.Therefore,in recent years,researchers have progressed in the study of algorithms and models for violence detection.In order to enhance practicability of a violence detection model in real life,the model must have a favorable generalization ability.However,improving the generalization ability of existing models is still a challenge.Different from traditional machine learning methods,a deep learning model requires a large amount of data to support its training process.However,the reality is that the scale and number of violence detection datasets are not satisfactory,resulting in poor generalization ability of a violence detection method based on deep learning model.Therefore,in order to improve the generalization ability of a deep learning based detection model,datasets of violence need to be further enriched and optimized due that the datasets are fundamental for a violence detection model.Furthermore,based on the rich data resource,researchers also need to improve the structure of the model.Firstly,aiming to deal with the above mentions problems,this paper proposes two self-built violence datasets.Then,based on the datasets,two deep learning-based network models for spatiotemporal feature extraction are proposed aiming to realize adaptive violence detection methods.The main work of this paper is as follows:(1)The article analyzes existing violence detection models from two research directions:machine learning and deep learning,and explain the characteristics of these models.(2)To overcome shortcoming of the current violence detection datasets,such as small amount of data,single scene and low definition,we propose two self-built violence datasets,named as BL-2314 and HK-10000.The video data in these two datasets are mainly web videos for inclusion,which are subsequently edited and formatted into two categories,i.e.,violent and non-violent behaviors.(3)A deep learning violence detection model based on spatiotemporal feature extraction— a heterogeneous two-stream violence detection model is proposed.The heterogeneous two-stream violence detection model uses two branches network with different structures to first extract features in temporal and spatial,then fuse the features extracted from these two branches,and finally obtain a judgment result.The experimental results show that the model has a favorable generalization ability.Moreover,the model also exhibits excellent performance on both large-scale datasets and small-scale datasets.(4)Although the previous model has a better generalization capability,it has a very complex structure and is not efficient to run.To overcome this drawback,another improved violence detection model based on deep learning spatiotemporal feature extraction,named as the two stream C3 D violence detection model,is proposed.The two-stream C3 D violence detection model uses two identical C3 D network to first extract temporal and spatial features,and then fuse their features,and finally obtain the judgment result.Through the experimental results,it can be seen that the two-stream C3 D violence detection model not only has a good generalization ability,but also runs faster and has a more concise model structure.