The traditional methods are easy to generate a large number of fake samples or data loss when classifying unbalanced data.Therefore,this paper proposes a novel DBSCAN(density-based spatial clustering of application wi...The traditional methods are easy to generate a large number of fake samples or data loss when classifying unbalanced data.Therefore,this paper proposes a novel DBSCAN(density-based spatial clustering of application with noise)for data clustering.The density-based DBSCAN clustering decomposition algorithm is applied to most classes of unbalanced data sets,which reduces the advantage of most class samples without data loss.The algorithm uses different distance measurements for disordered and ordered classification data,and assigns corresponding weights with average entropy.The experimental results show that the new algorithm has better clustering effect than other advanced clustering algorithms on both artificial and real data sets.展开更多
With the rapid development of related computer vision algorithms,the large-scale use of video surveillance systems has not only improved traffic safety,but also promoted the development of intelligent high-speed.Howev...With the rapid development of related computer vision algorithms,the large-scale use of video surveillance systems has not only improved traffic safety,but also promoted the development of intelligent high-speed.However,due to the complexity of the application scene,especially in the face of complex scene occlusion factors,the noise generated by the occlusion inevitably leads to the loss of the feature information of the identified person or object,which poses a great challenge to the existing pedestrian re-recognition algorithms.Therefore,this paper proposes a novel pedestrian re-recognition based on hybrid network.Feature extraction is carried out on four cooperative branches:local branch,global branch,global contrast pool branch and associated branch,and powerful diversity pedestrian feature expression ability is obtained.The network in this paper can be applied to different backbone networks.Through experimental comparison,the proposed algorithm has certain advantages compared with the latest methods,and the ablation experimental analysis further proves the effectiveness of the proposed network structure.展开更多
Surface EMG contains a lot of physiological information reflecting the intention of human movement.Gesture recognition by surface EMG has been widely concerned in the field of human-computer interaction and rehabilita...Surface EMG contains a lot of physiological information reflecting the intention of human movement.Gesture recognition by surface EMG has been widely concerned in the field of human-computer interaction and rehabilitation.At present,most studies on gesture recognition based on surface EMG signal are obtained by discrete separation method,ignoring continuous natural motion.A gesture recognition method of surface EMG based on improved long short-term memory network is proposed.sEMG sensors are rationally arranged according to physiological structure and muscle function.In this paper,the finger curvature is used to describe the gesture state,and the gesture at every moment can be represented by the set of different finger curvature,so as to realize continuous gesture recognition.Finally,the proposed gesture recognition model is tested on Ninapro(a large gesture recognition database).The results show that the proposed method can effectively improve the representation mining ability of surface EMG signal,and provide reference for deep learning modeling of human gesture recognition.展开更多
In this paper,abnormal target detection and location in video surveillance system are studied.In recent years,with the rapid development of network information technology,video surveillance technology has been widely ...In this paper,abnormal target detection and location in video surveillance system are studied.In recent years,with the rapid development of network information technology,video surveillance technology has been widely used,artificial anomaly detection methods have no way to meet the effective growth of video surveillance data,with 3D technology,face recognition technology,etc.,also promote the development of the field of computer vision,for the rapid analysis of a large number of video data to provide effective support.At present,abnormal target detection methods in video surveillance system mainly include the following two methods:One is to extract two-dimensional data features from video surveillance data,and effectively express video targets according to the extracted features.The information expressed mainly includes time information and spatial information.The second is to directly learn 3D space-time features for the module with motion information to detect the location of the abnormal target.Finally,the paper summarizes the full text and looks forward to the future development direction of video anomaly detection from three aspects:data set,method and evaluation index.展开更多
文摘The traditional methods are easy to generate a large number of fake samples or data loss when classifying unbalanced data.Therefore,this paper proposes a novel DBSCAN(density-based spatial clustering of application with noise)for data clustering.The density-based DBSCAN clustering decomposition algorithm is applied to most classes of unbalanced data sets,which reduces the advantage of most class samples without data loss.The algorithm uses different distance measurements for disordered and ordered classification data,and assigns corresponding weights with average entropy.The experimental results show that the new algorithm has better clustering effect than other advanced clustering algorithms on both artificial and real data sets.
基金supported by the National Natural Science Foundation of China(51622102,51571124,21421001)the 111 Project(B12015)+1 种基金the Natural Science Foundation of Tianjin(18ZXJMTG00040,16PTSYJC00030)the Fundamental Research Funds for the Central Universities~~
文摘With the rapid development of related computer vision algorithms,the large-scale use of video surveillance systems has not only improved traffic safety,but also promoted the development of intelligent high-speed.However,due to the complexity of the application scene,especially in the face of complex scene occlusion factors,the noise generated by the occlusion inevitably leads to the loss of the feature information of the identified person or object,which poses a great challenge to the existing pedestrian re-recognition algorithms.Therefore,this paper proposes a novel pedestrian re-recognition based on hybrid network.Feature extraction is carried out on four cooperative branches:local branch,global branch,global contrast pool branch and associated branch,and powerful diversity pedestrian feature expression ability is obtained.The network in this paper can be applied to different backbone networks.Through experimental comparison,the proposed algorithm has certain advantages compared with the latest methods,and the ablation experimental analysis further proves the effectiveness of the proposed network structure.
文摘Surface EMG contains a lot of physiological information reflecting the intention of human movement.Gesture recognition by surface EMG has been widely concerned in the field of human-computer interaction and rehabilitation.At present,most studies on gesture recognition based on surface EMG signal are obtained by discrete separation method,ignoring continuous natural motion.A gesture recognition method of surface EMG based on improved long short-term memory network is proposed.sEMG sensors are rationally arranged according to physiological structure and muscle function.In this paper,the finger curvature is used to describe the gesture state,and the gesture at every moment can be represented by the set of different finger curvature,so as to realize continuous gesture recognition.Finally,the proposed gesture recognition model is tested on Ninapro(a large gesture recognition database).The results show that the proposed method can effectively improve the representation mining ability of surface EMG signal,and provide reference for deep learning modeling of human gesture recognition.
文摘In this paper,abnormal target detection and location in video surveillance system are studied.In recent years,with the rapid development of network information technology,video surveillance technology has been widely used,artificial anomaly detection methods have no way to meet the effective growth of video surveillance data,with 3D technology,face recognition technology,etc.,also promote the development of the field of computer vision,for the rapid analysis of a large number of video data to provide effective support.At present,abnormal target detection methods in video surveillance system mainly include the following two methods:One is to extract two-dimensional data features from video surveillance data,and effectively express video targets according to the extracted features.The information expressed mainly includes time information and spatial information.The second is to directly learn 3D space-time features for the module with motion information to detect the location of the abnormal target.Finally,the paper summarizes the full text and looks forward to the future development direction of video anomaly detection from three aspects:data set,method and evaluation index.