一种基于深度强化学习的室内声学行为识别方法

对声学行为识别的研究目前主要依赖于特定用户的数据,且需要过滤异常值,导致较难获取可用于训练的数据集.提出了一种基于梅尔频谱图与Google AudioSet中提取的embedding的新策略,保证了模型的泛化能力,摆脱了依赖特定用户数据的限制.使用深度强化学习方法对11种常见室内行为进行识别,动态控制数据分布,解决数据不平衡问题.总体识别准确率达到87.5%,对每个行为的识别准确率均超过了83%.

An acoustic activity recognition based on deep reinforcement learning

Most of previous research normally relied on specific data and manual filtering of outliers for better performance.In this paper,a new strategy of activity recognition was proposed which was entirely free from the constraint of user data and guaranteed the generalization ability of model by usage of combined Mel spectrogram and embedding features extracted from video sound clips of Google AudioSet dataset.11 general domestic-related activities were recognized and evaluated based on deep reinforcement learning method,which dynamically controlled the distribution of data and resolved the data imbalance problem.The experimental test produced 87.5% overall accuracy and more than 83% accuracy for the 11 activities respectively.