Human action recognition is the key technique in many modern intelligent systems, such as video surveillance, sport motion analysis, virtual reality, and shopping behavior analysis. Human action recognition is a challenging task due to the sophisticated backgrounds and possible occlusions In this paper, we aim to recognize human actions from the video stream of sport scenes by intelligently fusing multimodel features. First of all, we leverage the background difference strategy and the histogram to extract human body contour from each video stream. This will be further processed by human morphology. Subsequently, we leverage the vectorization of human body contour and parts to divide each person into a set of sub-regions. Afterward, the key points are extracted from these sub-regions. We propose to fuse key points of human body, audio features, which are further fed into a support vector machine (SVM) to classify different human actions. Comprehensive experimental results have demonstrated the effectiveness of our designed recognition pipeline.

人体动作识别是许多现代智能系统中的关键技术，例如视频监视，运动动作分析，虚拟现实和购物行为分析。由于复杂的背景和可能存在的遮挡，人体动作识别是一项具有挑战性的任务。在本文中，我们旨在通过智能融合多模型特征从运动场景的视频流中识别人体动作。首先，我们利用背景差异策略和直方图从每个视频流中提取人体轮廓。这将通过人类形态进一步处理。随后，我们利用人体轮廓和部位的矢量化将每个人划分为一组子区域。然后，从这些子区域中提取关键点。我们建议融合人体关键点，音频功能，并将其进一步输入支持向量机（SVM），以对不同的人类行为进行分类。全面的实验结果证明了我们设计的识别管道的有效性。