A sequence models-based real-time multi-person action recognition method with monocular vision

In intelligent video surveillance under complex scenes, it is vital to identify the current actions of multi-target human bodies accurately and in real time. this paper, a real-time multi-person action recognition method with monocular vision proposed based on sequence models. Firstly, key points body skeleton are extracted by using OpenPose algorithm. Then, features constructed, including limb direction vector height-width ratio. The tracking then achieved Next, results matched features, model which includes spatial branch Deep neural networks temporal Bi-directional RNN long short-term memory networks. After pre-training, can be used recognize from stabilizer designed minimize false alarms. Finally, extensive evaluations JHMDB dataset validate effectiveness superiority approach.