Rehabilitation exercise is an important treatment method for stroke patients.This paper proposes a rehabilitation action recognition algorithm,Pose-AMGRU,which combines pose estimation with Gated Recurrent Unit(GRU) in order to improve the accuracy and real-time performance of rehabilitation action recognition,and thus assist patients in in-home long-term rehabilitation training.The algorithm uses OpenPose pose estimation method to extract the skeleton joints from video frames,and the pose data is preprocessed to obtain the key action features that represent body movement.Then a GRU network with three-layer time series features is constructed by using the attention mechanism to realize rehabilitation action classification.Experimental results on KTH dataset and rehabilitation action dataset show that the proposed algorithm increases the recognition accuracy to 98.14% and 100%,and its running speed on GTX1060 reaches 14.23 frame/s,which demonstrates its excellent recognition accuracy and real-time performance.

为了进一步提高居家监护场景下人体动作识别的可靠性与实时性,更好地辅助出院后的卒中患者进行康复训练,提出一种基于单目视觉的在线人体动作识别算法.融合姿态估计OpenPose与最近邻匹配算法对监控视频流中的目标人体生成动作序列.通过滑动窗口选取原始姿态特征并对其预处理转化为鲁棒性特征,输入到多层LSTM长短时记忆网络中进行康复动作识别.实验结果表明,该方法对活动背景、人体穿着、无关人员的干扰等具有较强的适应能力,能够在线识别连续的康复动作且准确率达90.66％,在居家康复训练场景中有一定的应用价值.

The irrelevant background information in the personalized image is easy to be quantified into the same word as the main target, and the quantization process will inevitably cause the loss of a lot of visual information. This phenomenon will seriously reduce the quality of the generated theme when the personalized image content is complex. This paper proposes a Multi-Source Big Data Fusion Annotation (MSBDFA) model. The model obtains similar personalized images by analyzing the relevant multi-source information of the personalized images, and uses the annotations of the similar personalized images to label the personalized images. For the personalized images with complex background visual information, the personalized image retrieval based on complete information modeling uses the high-dimensional Gaussian distribution to directly model the continuous visual features of the personalized images, and uses the two-level spectral clustering algorithm to distribute the regional topics, so as to embed the complete local information contained in the visual features into the global features of the personalized image. Therefore, this method can completely retain visual information during the modeling process, so that the targets buried in the complex background can be better classified. The experimental results on the standard database show that the method proposed in this paper can generate high-quality personalized image subjects in complex scenes and has good retrieval performance.

The action recognition of human rehabilitation movement in the home scene plays a positive role in promoting the rehabilitation process of patients, we present an efficient approach for real-time continuous human rehabilitation action recognition using OpenPose and FCN. The proposed method first fuses OpenPose with Kalman filter to track human targets and generate the 2D poses action sequences from the RGB videos stream. Then we extract the segmented action sequence by sliding the window and convert the rectangular coordinates to relative coordinates from each frame of the human skeleton. We design a 1D fully convolutional network to extract spatial-temporal characteristics and classify actions. The experimental results show that the method has strong adaptability to the interference of activity background, human body wearing and irrelevant personnel, and it can identify continuous rehabilitation actions online with an accuracy rate of 85.6%. At the same time, the model running speed on GTX1060 is up to 18.14 FPS which has certain application value in the home rehabilitation training scene.