<p>Embedding the behavior recognition network into the edge-side video surveillance device can solve the problems of high latency, high computational resource consumption, and high cost of the cloud placement scheme. However, in placing the network on the edge side, Ignoring the relationship between network complexity and embedded devices will lead to the embedded devices failing to meet the identification network’s requirements and the identification network’s performance not meeting the scene’s requirements. To address the above problem, we tested the recognition time of different recognition networks, explored the relationship between the recognition time of the network and the amount of network computation and parameters, used fitting tools to solve the relationship between the three variables, and combined the solved model with different hardware performance to modify it, propose a multiparameter fusion model of a behavior recognition network and realize the prediction of the recognition time of the recognition network. The model predicted a recognition time error of within 5 % in tests on UCF-101 dataset and production scene behavior dataset. Taking the design of the dangerous behavior identification network of CRRC Zhuzhou Locomotive Factory as an example, We use this model to predict and optimize the recognition network’s recognition time, modify the recognition network’s structure, and reduce the recognition time of the network.</p>