The paper considers the problem of choosing an optimal video monitoring mode when using neural network models as a recognizer when different models are more effective on a video stream at different times. Video monitoring tasks are different while the conditions for obtaining data are different, which can be expressed in the recognition complexity concept. Evaluation of the recognition complexity in monitoring allows saving computing resources, thereby reducing the cost of implementation and use. After evaluating the average complexity of recognition, it is possible to choose the optimal recognition mode in terms of speed and reliability during post-processing, when time for it is limited.The paper shows the problem solution in the task of two type object detection using YOLOv5 models, when the video stream must be processed in real time with a minimum delay when the result is returned after each frame. The metrics used in the object detection are analyzed in terms of a possibility of assessing the reliability of the results when there is no final information about an object. There is a chosen efficiency criterion based on the sum of the F1-score and the cost of computing resources, which makes it possible to evaluate the model effectiveness for specific objects. The paper shows the dependence of the efficiency criterion on the F1-score for two models. There are the results of testing two models and a dynamic mode based on choosing an appropriate model depending on the input object. The paper describes the limitations of the approach, which can be used only for streaming recognition, when the images received for recognition are only slightly different from the previous ones. in the end, there is a conclusion about the approach applicability for a number of problems in accordance with the restrictions.