【摘要】 针对目前矿工行为数据集构建不全面、行为识别实时性较差、对相似行为的细粒性识别精度较低等问题,提出了一种端到端的自主学习行为特征并实现行为分类的识别方法。首先,对原始矿工行为视频进行特征提取,生成用来描述时间特征的光流图以及可以描述空间特征的三原色(RGB)图像,使用双流网络对提取的特征进行学习并得到行为分类结果;然后,引入量子遗传算法对双流网络进行改进,对网络中待训练参数进行量子编码,将双流网络在测试集上的代价函数值作为适应度函数。采用量子交叉、量子门旋转实现种群个体的进化。构建了包含50种矿工行为的数据集,在该数据集上利用双流法进行行为识别。研究结果表明:使用量子遗传算法优化后的3种双流网络的识别准确率,比优化前分别提升了1.01%,0.87%和0.32%。通过与其他矿工行为识别算法进行对比,本文所提方法在两种数据集上识别率分别达到90.36%和72.29%,均优于其他几种识别算法,准确率最大差距达到22.36%,证明了本文所提方法的有效性。更多还原

【Abstract】 In view of incomplete construction of current miner behavior datasets,poor real-time behavior recognition and low accuracy of fine-grained recognition of similar behavior,an end-to-end recognition method for self-learning behavior feature and behavior classification was proposed. Firstly,feature extraction was performed on the original miner behavior video to generate optical flow graphs describing temporal features and three-primary colours( RGB) images describing spatial features. A two-stream network was used to learn the extracted feature and obtain behavior classification results. Then the quantum genetic algorithm was introduced to improve the two-stream network,and the parameters to be trained in the network were quantum-coded. The cost function of the two-stream network on the test-set was used as the fitness function,and quantum cross and quantum gate rotation were used to realize the evolution of the individual population. A dataset containing 50 kinds of miner behaviors was constructed. On this dataset,two-flow method was used for behavior recognition.The results show that the recognition rate of three dual-flow networks optimized by the quantum genetic algorithm are improved by 1. 01%,0. 87% and 0. 32%,respectively. Moreover,compared with other miner behavior recognition algorithms,the recognition rate of the proposed method in this paper reaches to 90.36%and 72.29%,respectively on the two datasets,both of which are superior to other miners behavior recognition algorithms. The maximum accuracy difference reaches to 22. 36%,which proves the effectiveness of the proposed method.更多还原