This study was undertaken to explore 18 time domain (TD) and time-frequency domain (TFD) feature configurations to determine the most discriminative feature sets for classification. Features were extracted from the surface electromyography (sEMG) signal of 17 hand and wrist movements and used to perform a series of classification trials with the random forest classifier. Movement datasets for 11 intact subjects and 9 amputees from the NinaPro online database repository were used. The aim was to identify any optimum configurations that combined features from both domains and whether there was consistency across subject type for any standout features. This work built on our previous research to incorporate the TFD, using a Discrete Wavelet Transform with a Daubechies wavelet. Findings report configurations containing the same features combined from both domains perform best across subject type (TD: root mean square (RMS), waveform length, and slope sign changes; TFD: RMS, standard deviation, and energy). These mixed-domain configurations can yield optimal performance (intact subjects: 90.98%; amputee subjects: 75.16%), but with only limited improvement on single-domain configurations. This suggests there is limited scope in attempting to build a single absolute feature configuration and more focus should be put on enhancing the classification methodology for adaptivity and robustness under actual operating conditions.
{"title":"Effectiveness of Surface Electromyography in Pattern Classification for Upper Limb Amputees","authors":"Carl Peter Robinson, Baihua Li, Q. Meng, M. Pain","doi":"10.1145/3268866.3268889","DOIUrl":"https://doi.org/10.1145/3268866.3268889","url":null,"abstract":"This study was undertaken to explore 18 time domain (TD) and time-frequency domain (TFD) feature configurations to determine the most discriminative feature sets for classification. Features were extracted from the surface electromyography (sEMG) signal of 17 hand and wrist movements and used to perform a series of classification trials with the random forest classifier. Movement datasets for 11 intact subjects and 9 amputees from the NinaPro online database repository were used. The aim was to identify any optimum configurations that combined features from both domains and whether there was consistency across subject type for any standout features. This work built on our previous research to incorporate the TFD, using a Discrete Wavelet Transform with a Daubechies wavelet. Findings report configurations containing the same features combined from both domains perform best across subject type (TD: root mean square (RMS), waveform length, and slope sign changes; TFD: RMS, standard deviation, and energy). These mixed-domain configurations can yield optimal performance (intact subjects: 90.98%; amputee subjects: 75.16%), but with only limited improvement on single-domain configurations. This suggests there is limited scope in attempting to build a single absolute feature configuration and more focus should be put on enhancing the classification methodology for adaptivity and robustness under actual operating conditions.","PeriodicalId":285628,"journal":{"name":"Proceedings of the 2018 International Conference on Artificial Intelligence and Pattern Recognition","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130150331","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pattern recognition of time-series signals for movement and gesture analysis plays an important role in many fields as diverse as healthcare, astronomy, industry and entertainment. As a new technique in recent years, Deep Learning (DL) has made tremendous progress in computer vision and Natural Language Processing (NLP), but largely unexplored on its performance for movement and gesture recognition from noisy multi-channel sensor signals. To tackle this problem, this study was undertaken to classify diverse movements and gestures using four developed DL models: a 1-D Convolutional neural network (1-D CNN), a Recurrent neural network model with Long Short Term Memory (LSTM), a basic hybrid model containing one convolutional layer and one recurrent layer (C-RNN), and an advanced hybrid model containing three convolutional layers and three recurrent layers (3+3 C-RNN). The models will be applied on three different databases (DB) where the performances of models were compared. DB1 is the HCL dataset which includes 6 human daily activities of 30 subjects based on accelerometer and gyroscope signals. DB2 and DB3 are both based on the surface electromyography (sEMG) signal for 17 diverse movements. The evaluation and discussion for the improvements and limitations of the models were made according to the result.
{"title":"Movement and Gesture Recognition Using Deep Learning and Wearable-sensor Technology","authors":"Baao Xie, Baihua Li, A. Harland","doi":"10.1145/3268866.3268890","DOIUrl":"https://doi.org/10.1145/3268866.3268890","url":null,"abstract":"Pattern recognition of time-series signals for movement and gesture analysis plays an important role in many fields as diverse as healthcare, astronomy, industry and entertainment. As a new technique in recent years, Deep Learning (DL) has made tremendous progress in computer vision and Natural Language Processing (NLP), but largely unexplored on its performance for movement and gesture recognition from noisy multi-channel sensor signals. To tackle this problem, this study was undertaken to classify diverse movements and gestures using four developed DL models: a 1-D Convolutional neural network (1-D CNN), a Recurrent neural network model with Long Short Term Memory (LSTM), a basic hybrid model containing one convolutional layer and one recurrent layer (C-RNN), and an advanced hybrid model containing three convolutional layers and three recurrent layers (3+3 C-RNN). The models will be applied on three different databases (DB) where the performances of models were compared. DB1 is the HCL dataset which includes 6 human daily activities of 30 subjects based on accelerometer and gyroscope signals. DB2 and DB3 are both based on the surface electromyography (sEMG) signal for 17 diverse movements. The evaluation and discussion for the improvements and limitations of the models were made according to the result.","PeriodicalId":285628,"journal":{"name":"Proceedings of the 2018 International Conference on Artificial Intelligence and Pattern Recognition","volume":"22 6","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"120992363","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Over the past few decades, Optical Character Recognition (OCR), particularly handwriting recognition, has received much attention. Handwritten Digits Recognition (HDR) means, receive and comprehend handwriting inputs from different sources for example pictures, touch screens, paper documents, and other devices. The field of HDR has witnessed rapid progress owing to the concurrent availability of cheap and well-assembled computers, advancements in learning algorithms, and availability of large databases. In recent years, HDR has received much attention due to ambiguity in learning methods. The aim of the current study was to explore the potential of Deeplearnig4j (DL4J) framework for HDR. DL4J offers the most appropriate framework for the identification of handwritten digits. To execute the task of HDR, Convolutional Neural Network (CNN) is implemented. This study measures the strength and productivity of DL4J for the aforementioned tasks of recognition and attempts to upgrade the procedure. Results obtained shows significant improvement in the recognition rates of hand-typed digits. Though the accuracy and error rates obtained through our proposed system (CNN-DL4J) show variations, on average the accuracy rate remained at 97 %. The aim of the proposed endeavor was to make the path towards digitalization clearer. Though the purpose was only to identify the digits, we can extend it to deal with digits having different sizes, different languages (Urdu, Arabic, Persian), letters, and the task of detecting multi-digit person's handwriting. Hence, it could reduce the typing need to an extent that people will be able to convert their handwritten materials into digital form in one click on its picture. Altogether, this investigation combines CNN with the DL4J framework and took MNIST as a standard dataset to accomplish the task of digit recognition. In addition, the test framework can be assessed in the future for the prospects of image classification and such other pattern recognition tasks.
{"title":"Handwritten Digits Recognition Based on Deep Learning4j","authors":"Zareen Sakhawat, Saqib Ali, Li Hongzhi","doi":"10.1145/3268866.3268888","DOIUrl":"https://doi.org/10.1145/3268866.3268888","url":null,"abstract":"Over the past few decades, Optical Character Recognition (OCR), particularly handwriting recognition, has received much attention. Handwritten Digits Recognition (HDR) means, receive and comprehend handwriting inputs from different sources for example pictures, touch screens, paper documents, and other devices. The field of HDR has witnessed rapid progress owing to the concurrent availability of cheap and well-assembled computers, advancements in learning algorithms, and availability of large databases. In recent years, HDR has received much attention due to ambiguity in learning methods. The aim of the current study was to explore the potential of Deeplearnig4j (DL4J) framework for HDR. DL4J offers the most appropriate framework for the identification of handwritten digits. To execute the task of HDR, Convolutional Neural Network (CNN) is implemented. This study measures the strength and productivity of DL4J for the aforementioned tasks of recognition and attempts to upgrade the procedure. Results obtained shows significant improvement in the recognition rates of hand-typed digits. Though the accuracy and error rates obtained through our proposed system (CNN-DL4J) show variations, on average the accuracy rate remained at 97 %. The aim of the proposed endeavor was to make the path towards digitalization clearer. Though the purpose was only to identify the digits, we can extend it to deal with digits having different sizes, different languages (Urdu, Arabic, Persian), letters, and the task of detecting multi-digit person's handwriting. Hence, it could reduce the typing need to an extent that people will be able to convert their handwritten materials into digital form in one click on its picture. Altogether, this investigation combines CNN with the DL4J framework and took MNIST as a standard dataset to accomplish the task of digit recognition. In addition, the test framework can be assessed in the future for the prospects of image classification and such other pattern recognition tasks.","PeriodicalId":285628,"journal":{"name":"Proceedings of the 2018 International Conference on Artificial Intelligence and Pattern Recognition","volume":"9 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131836187","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pedestrian protection system plays an important role in perceptual system of unmanned vehicles and Advanced Drive Assistant System. In order to get more details information about surrounding objects, perceptual system of such kind intelligence system is usually equipped with different sensors, so technology to fuse information of heterogeneous sensors is very important. This paper proposed a novel way to fuse the information of radar and image of camera to realize pedestrian detection and acquire its dynamic information. Contribution of this paper are as following First, a new intra-frame cluster algorithm and an inter-frame tracking algorithm are put forward to extract valid target signal from original radar data with noise. Second, to realize radar-vision data space alignment, least square algorithm is used to get the coordinate transformation matrix. Then with the aid of projections of radar points, a flexible strategy to generate region of interest (ROI) is put forward. Furthermore, to further accelerate detection, an improved fast object estimation algorithm is proposed to acquire a more accurate potential target area based on ROI. At last, histogram of gradient (HOG) features of potential area are extracted and support vector machine is used to judge whether it's a pedestrian. The proposed approach is verified through real experimental examples and the trial results show this method is feasible and effective.
{"title":"Pedestrian Detection Based on Fusion of Millimeter Wave Radar and Vision","authors":"Xiao Guo, Jinsong Du, Jie Ying Gao, Wei Wang","doi":"10.1145/3268866.3268868","DOIUrl":"https://doi.org/10.1145/3268866.3268868","url":null,"abstract":"Pedestrian protection system plays an important role in perceptual system of unmanned vehicles and Advanced Drive Assistant System. In order to get more details information about surrounding objects, perceptual system of such kind intelligence system is usually equipped with different sensors, so technology to fuse information of heterogeneous sensors is very important. This paper proposed a novel way to fuse the information of radar and image of camera to realize pedestrian detection and acquire its dynamic information. Contribution of this paper are as following First, a new intra-frame cluster algorithm and an inter-frame tracking algorithm are put forward to extract valid target signal from original radar data with noise. Second, to realize radar-vision data space alignment, least square algorithm is used to get the coordinate transformation matrix. Then with the aid of projections of radar points, a flexible strategy to generate region of interest (ROI) is put forward. Furthermore, to further accelerate detection, an improved fast object estimation algorithm is proposed to acquire a more accurate potential target area based on ROI. At last, histogram of gradient (HOG) features of potential area are extracted and support vector machine is used to judge whether it's a pedestrian. The proposed approach is verified through real experimental examples and the trial results show this method is feasible and effective.","PeriodicalId":285628,"journal":{"name":"Proceedings of the 2018 International Conference on Artificial Intelligence and Pattern Recognition","volume":"71 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-08-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132336998","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Proceedings of the 2018 International Conference on Artificial Intelligence and Pattern Recognition","authors":"","doi":"10.1145/3268866","DOIUrl":"https://doi.org/10.1145/3268866","url":null,"abstract":"","PeriodicalId":285628,"journal":{"name":"Proceedings of the 2018 International Conference on Artificial Intelligence and Pattern Recognition","volume":"182 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1900-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133716593","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：