基于改进的深度神经网络的人体动作识别模型

针对现有人体动作识别方法需输入固定长度的视频段、未充分利用时空信息等问题,提出一种基于时空金字塔和注意力机制相结合的深度神经网络模型,将包含时空金字塔的3D-CNN和添加时空注意力机制的LSTM模型相结合,实现了对视频段的多尺度处理和对动作的复杂时空信息的充分利用。以RGB图像和光流场作为空域和时域的输入,以融合金字塔池化层的运动和外观特征后的融合特征作为融合域的输入,最后采用决策融合策略获得最终动作识别结果。在UCF101和HMDB51数据集上进行实验,分别取得了94.2%和70.5%的识别准确率。实验结果表明,改进的网络模型在基于视频的人体动作识别任务上获得了较高的识别准确率。     

基于时空域深度神经网络的野火视频烟雾检测

针对目前的烟雾检测算法主要基于单一特征或烟雾的多个动静态特征的融合导致检测精度低的问题,提出一种使用卷积神经网络和循环神经网络组合的视频烟雾检测框架来捕获烟雾在空间域和时间域中的特征信息。利用空间流网络部分对运动区域自动提取特征后进行初步的空域的判别;在将空域判断为有烟的基础上进一步通过时间流网络和循环神经网络部分累积一组连续帧之间的运动信息以区分烟雾和非烟雾区域。与现有的使用深度卷积神经网络模型进行对比实验,实验结果表明,该方法具有较高的分类检测准确率。在多个视频场景中进行测试,验证了该算法的有效性。     

基于运动筛选和3D卷积的视频早期烟雾检测

To solve the problem of high false alarm and high missed detection in the complex environment of early smoke detection based on video, a method based on motion extraction of suspected areas is proposed and a multi-scale 3D convolutional neural network with input of 6 frames（6M3DC） is designed for video smoke detection. Firstly, the motion regions are obtained through the background difference model after average filtering and the positions of the block in which the motion regions are located are calculated, and then the motion blocks are extracted by color judgment and mean HASH algorithm and the nonconforming blocks are updated to the background image. Finally, by combining the suspected blocks of the same region of 6 consecutive frames as the input for the 3D convolutional neural network for detection, blocks detected as smoke are marked and non-smoke blocks are updated to the background image. The experimental results show that the algorithm is adaptive to slow moving smoke and can detect smoke in complex environment.