基于视角不变全局特征表达的人体行为检测与识别算法研究

This project tries to use view-invariant world representations to detect and recognize the human behavior in videos, so as to solve the challenges that the traditional recognition methods have low accuracy and are difficult to apply to the complex scenes. Faster-RCNN is used to identify the dynamic characteristics in the spatial and time domains based on multi-frame learning. Event detection is performed by the maximum margin event detection methods, especially for a large number of events, the detection speed is good. Based on multi-angle guidance, a multi-level fine-grained detection method for bodies’ key points is proposed to solve the problem of the detail information in the real data and target occlusion, and the accuracy of recognition is improved. Using the geometric scene data, the pedestrian appearance model is generated without the mark real data by combining with the effective difference learning method, and the applicable scene is expanded. We can utilize the multi-segment SVM, the spatial component and the semantic component to infer the state of the target, predict the target activity and find the video segment. We are planning to verify the practicability and efficiency of the solutions in real scenes by doing experiments in some related experiments. Its scientific essence is to introduce convolution neural network and other deep learning theories into video behavior recognition, and further to identify the movement characteristics of human behavior in videos. The research results provide a theoretical support and application reference for video behavior detection and recognition.

本项目尝试利用视角不变全局特征表达对视频监测中人体行为进行检测与识别，从而解决传统视频识别方法准确度低、难以适用复杂场景等问题。基于多帧学习，利用Faster-RCNN识别空间域和时域中的动态特征，结合最大余量事件检测方法，进行行为和事件检测，对复杂场景有良好的检测速度。基于多视角引导，提出一种多层次细粒度的人体关键点检测方法，解决真实数据中细节信息复杂以及目标被遮挡等问题，提高准确性；利用几何场景数据，结合有效的差别学习方法，在无标记真实数据的情况下生成行人外观模型，拓展了适用场景。通过多段SVM，空间分量和语义分量来推断目标的状态，预测目标活动，寻找视频中的辨别段。项目经通过相关实验环境验证所提方法的有效性。其科学实质是在视频行为识别中引入卷积神经网络等深度学习和机器学习理论，进一步识别视频中人体行为的运动特征，研究成果为视频行为检测和识别提供了理论支撑和应用参考。

本项目尝试利用视角不变全局特征表达对视频监测中人体行为进行检测与识别，从而解决传统视频识别方法准确度低、难以适用复杂场景等问题。基于多帧学习，利用Faster-RCNN识别空间域和时域中的动态特征，结合最大余量事件检测方法，进行行为和事件检测，对复杂场景有良好的检测速度。基于多视角引导，提出一种多层次细粒度的人体关键点检测方法，解决真实数据中细节信息复杂以及目标被遮挡等问题，提高准确性；利用几何场景数据，结合有效的差别学习方法，在无标记真实数据的情况下生成行人外观模型，拓展了适用场景。通过多段SVM，空间分量和语义分量来推断目标的状态，预测目标活动，寻找视频中的辨别段。项目经通过相关实验环境验证所提方法的有效性。其科学实质是在视频行为识别中引入卷积神经网络等深度学习和机器学习理论，进一步识别视频中人体行为的运动特征，研究成果为视频行为检测和识别提供了理论支撑和应用参考。