Application of neural network algorithms information processing in autonomous video monitoring systems

The article is devoted to the problem of processing video information using neural networks in autonomous systems with limited computing resources. The purpose of the study. The purpose of the ar¬ticle is to study algorithms for processing video information in autonomous systems with limited computing resources based on two-stage data processing. At the first stage, the contours of the observed objects are determined using traditional methods of identifying key features. The second stage involves processing using neural network algorithms aimed at classifying and recognizing objects. Due to the primary processing of the video stream, the amount of information processed by the neural network is significantly reduced. Materials and methods. The article offers a variant of the algorithm for primary processing of video information based on methods for determining the contours of objects. Boundary allocation methods based on spatial filtering can be applied to both stationary and mobile objects. Secondary processing is based on the use of a convolutional neural network, the coefficients of which were quantized to reduce the computational load. To train the neural network, two datasets were generated for training, navigation, and testing. The first dataset consisted of initial images with an object, the second contained an ideal contour of objects (masks). A comparison of two neural networks trained on the source images and on contour masks of the images is carried out. The features of the implementation of neural networks on FPGAs are considered and a block diagram of the calculations is proposed. The possibility of iterative calculations with repeated use of computing resources at each iteration (neural network layer) is shown. Results. The results of preliminary studies on neural network training and quantization of its coefficients are presented, and the specifics of hardware implementation of neural networks based on programmable logic integrated circuits are considered. Conclusion. The two-stage approach to video information processing demonstrates high efficiency in autonomous systems where there are limitations on energy consumption and computing resources. The article may be useful to developers of autonomous video monitoring systems, detection and tracking of ground-based and air-based facilities.