System for detecting dynamic objects on video sequence frames

The paper considers the system of dynamic object detection and visualization of its results. A brief review of existing approaches to video monitoring data analysis is given, using smoke cloud detection as an example. The research considers the algorithm of object detection, which is based on the implementation of the EfficientDet-D1 model. Authors propose methods and algorithms of pre-processing, video predictions clustering and filtering detected objects by hybrid architecture of the neural network. Those methods and algorithms were consistently implemented to improve the efficiency of the neural network. The pre-processing algorithm allows to select dynamic features on the frame. The idea behind the post-processing algorithm is to combine the results of sequential detections of the characteristic features on each frame, in particular the smoke cloud features. The method of detected features filtering is implemented by an ensemble of recurrent and convolutional neural networks. The results of the system on the test sample: Precision - 98%, Recall 97%, Accuracy - 98%.