边缘智能是将AI技术应用于边缘嵌入式设备，提供一种智能化计算新范式，被广泛应用于物联网系统。智能摄像机是具有代表性的边缘产品，能够为智能家居、智能交通和智能监控提供低延迟的视频处理能力。由于摄像机计算资源的有限性，传统的行为识别模型难以在本地托管且及时地完成计算任务。本文设计了一种基于深度可分离卷积的长短时记忆学习模型-DSconv-LSTM，可以快速识别视频流中的目标行为。DSconv-LSTM使用深度可分离卷积来处理卷积LSTM学习单元中四个门的时空数据，从而大大降低了模型的复杂性。最后，利用两个公共视频数据集对DSconv-LSTM进行了评估。实验结果表明，DSconv-LSTM提升了模型的收敛性，大大减小了行为识别模型尺寸，加快了推理速度。

Edge computing provides an innovative construction paradigm for intelligent services on edge embedded devices with AI techniques, which has been used to improve the intelligence of IoT applications. Smart camera is one of representative intelligent edge products to be able to provide video-processing services for smart home, intelligent transportation and intelligent monitoring. Due to the resource constraint of cameras, some complex services, e.g., action recognition, are hard to be hosted locally to complete the computational tasks timely and precisely. In this paper, we design a lightweight model DSconv-LSTM based on depthwise separable convolution long short term memory learning unit for recognizing human behaviors locally through camera’s video streaming. The DSconv-LSTM uses depthwise separable convolution operation to handle the spatio-temporal data of four gates in convolution LSTM learning unit whereby complexity of recognition model is reduced greatly. Finally, two public video datasets of human behavior are used to test the DSconv-LSTM. The experimental result shows that the DSconv-LSTM improves the learning property on convergence, reduces the model size of action recognition greatly, and shortens the interference time of action recognition.