Search

Search Results

• IMPROVING THE ROBUSTNESS OF SIAMESE NETWORK DETECTION WITH LIMITED ANNOTATION THROUGH HARD EXAMPLE MINING

  S. Е. Babin , P.А. Gessen , V.А. Pavlova , V.А. Tupikov

  2026-04-29

  Abstract ▼

  The aim of this study is to develop and experimentally evaluate a method for improving the quality of marine vessel detection based on the Attention RPN architecture under conditions of limited training data, with the integration of inter-class hard negative mining. To achieve this goal, the following were implemented: model adaptation in a fine-tuning mode based on five examples, creation of a specialized dataset consisting of 9 vessel categories (637 training and 175 validation images), and an algorithm for selecting hard negative pairs using cosine similarity. The methodological foundation includes deep learning approaches for object detection based on Siamese neural networks, analysis of feature distributions in embedding space, and algorithms for generating hard negative samples by computing inter-class distances using the cosine metric. Experimental evaluation was carried out on standard few-shot object detection benchmarks. The integration of cross-class hard negative mining reduced the proportion of false positive detections by approximately 18% and increased the mean Average Precision (mAP) by about 6% on average compared to the baseline model without hard negative mining. The greatest improvement (up to 9.1%) was observed for classes with large target objects. The results demonstrate that targeted selection of inter-class hard negative examples significantly improves the robustness of metric-based detectors under conditions of strong class imbalance and high intra-class similarity. The practical significance of the work lies in the applicability of the proposed approach for deploying video surveillance systems in tasks where large training datasets and extensive annotation are unavailable

• OBJECT DETECTION ALGORITHM FOR OPTOELECTRONIC SYSTEMS WITH ONLINE LEARNING

  V.A. Tupikov, V.A. Pavlova, V.A. Bondarenko, M.V. Sozinova, P.A. Gessen

  2021-04-04

  Abstract ▼

  In order to create a new algorithm for automatic detection of objects with real-time training, a
  study of the world scientific groundwork in the field of general-purpose automatic tracking with the
  ability to recognize a tracked object with the potential for application in embedded computing systems
  of optoelectronic systems of promising robotic complexes was carried out. Based on the conducted
  research, methods and approaches were selected and tested that allow, with the greatest accuracy,
  while maintaining high computational efficiency, to provide training of classifiers on the fly
  (online learning) without a priori knowledge of the type of tracking object and to ensure the subsequent
  detection of the original object in the event of its short-term loss. Such methods include a histogram
  of oriented gradients – a descriptor of key features based on the analysis of the distribution of
  the brightness gradients of the object image. Its use allows you to reduce the amount of information
  used without losing key data about the object and to increase the speed of image processing. The
  article substantiates the choice of one of the real-time classification algorithms that allows solving
  the problem of binary classification – the support vector machine. Due to the high speed of data processing
  and the need for a small amount of initial training data to construct a separating hyperplane,
  on the basis of which the classification of objects is done, this method is chosen as the most suitable
  for solving the problem. For online training, a modification of the support vector machine method
  was chosen, which implements stochastic gradient descent at each step of the algorithm – Pegasos.
  The authors of the study carried out the development and semi-natural modeling of the selected algorithm,
  evaluated the effectiveness of its work in the tasks of detecting an object of interest in real time
  with preliminary online training in the process of tracking the object. The developed algorithm has
  shown high efficiency in solving the problem and is planned to be implemented as part of a special
  software for optoelectronic systems of advanced robotic systems. In the conclusion, proposals are
  presented to further improve the accuracy and probability of the object detection by the developed
  algorithm, as well as for improving its performance by optimizing calculations.

• CORRELATIONAL SUPPORT ALGORITHM WITH REAL-TIME LEARNING

  V. А. Tupikov, V. А. Pavlova, А.Y. Gagarina, P. А. Gessen, А.I. Lizin, М. V. Sozinova

  2022-04-21

  Abstract ▼

  In order to develop a stable algorithm for automatic detection and tracking of nondeterministic
  objects with real-time learning for embedded computing systems with optoelectronic
  devices, within the framework of this work, a study and analysis of the existing world scientific and
  technical experience in the field of automatic tracking algorithms for general purposes was carried
  out. The article shows that the most stable modern automatic tracking algorithms are a system
  that makes a decision about the current position, size and other parameters of the tracked
  image based on the model being trained. The authors of the study identified the most effective of
  the applied basic algorithms suitable for use in embedded computing systems of robotic complexes,
  and developed a new algorithm for automatic detection and maintenance of non-deterministic
  objects. A semi-natural testing of the developed algorithm was carried out and its effectiveness
  was evaluated in solving problems not only of automatic tracking of objects, but also problems of
  automatic detection of objects using several reference images. In conclusion, proposals are presented
  for further improving the accuracy of the developed algorithm and for its optimization and
  implementation in the special software of on-board computer systems of aircraft.

• RECURSIVE ANALYSIS ALGORITHM AND RESTORATION OF CONTOURS IN NAVIGATION AND GUIDANCE SYSTEMS

  V.А. Tupikov, V.А. Pavlova, А.I. Lizin, P.А. Gessen, V.D. Saenko

  2024-05-28

  Abstract ▼

  In order to develop an object detection algorithm for embedded computing systems of opticalelectronic
  complexes, an analysis of the existing world scientific and technical experience was carried out,
  aimed at improving the process of identifying contours. Based on the analysis, the authors of the article
  developed a new method for correcting contour images. This method implements an approach that allows
  you to merge broken contours and apply filtering based on various parameters for optimal contour analysis.
  The first step of the algorithm is to apply blur to the image, followed by the application of the Kenny
  edge detection algorithm. Then the contours are thinned and the contour image is filtered to remove the
  weakest contours. The next steps are the creation and processing of each individual contour, as well as
  filtering outliers. The final stage is to connect and search for inflection points of the contour. The work
  highlights both the advantages and disadvantages of classical edge extraction methods in the context of
  their use in object detection algorithms. The authors of the study analyzed two classical morphological
  operators - dilatation and erosion, as well as the existing basic variations of their use, such as opening
  and closing, as methods for combining contours. As a result of a comparative analysis of the results of the
  work of morphological operators of dilatation and erosion, as well as the main variations of their application,
  with a recursive algorithm for analyzing and restoring contours, the advantage of the latter in terms
  of preserving the integrity of the morphological characteristics of objects was revealed. The authors also
  proposed ideas for further development of a recursive algorithm for analysis and restoration of contours,
  as well as its further application in problems of detecting objects in images.

• HYBRID ALGORITHM OF AUTOMATIC TRACKING FOR EMBEDDED COMPUTERS OF OPTOELECTRONIC NAVIGATION AND GUIDANCE SYSTEMS

  V. А. Tupikov, V. А. Pavlova, А.I. Lizin, P.А. Gessen, V.D. Saenko

  2024-04-16

  Abstract ▼

  The authors of the work carried out research in the field of technical vision systems, as well
  as approaches to solving problems of detecting and tracking objects of interest without a priori
  knowledge of their type, taking into account the target platform in the form of an embedded optoelectronic
  system computer. Based on the data obtained, the sphere was analyzed and a new hybrid
  maintenance algorithm for embedded systems was proposed. It is based on a combination of several
  types of maintenance algorithms, with one of them as a priority, providing the main work, and
  several auxiliary ones to stabilize and expand the functionality of the priority one. They are connected
  by an external processing cycle, which, based on a consensus decision of internal algorithms,
  independently decides on the position of the target object in the frame and stores auxiliary
  information to ensure the correct operation of the entire algorithm, as well as responsible for making
  a decision on the re-detection of the target. The authors propose two possible implementations
  of this approach, used depending on the power of available computing resources. A variant of the
  algorithm has been implemented for the available computing power, and its semi-natural tests
  have been carried out based on real video sequences. They represent different backgrounds and
  different structural objects of interest with different dynamics of change over time. The evaluation
  of the results of the proposed algorithm in the tasks of detecting and tracking an object of interest
  in real time on the presented videos using a software package for automating testing of detection
  and tracking algorithms has been carried out. As a result, the algorithm showed high efficiency in
  the tasks set, improving the accuracy of tracking, in comparison with internal algorithms that
  worked separately, by adding rotary and scale invariances, and also significantly increased the
  ability to re-detect an object after its loss. In conclusion, the authors present proposals for the
  further development and implementation of optoelectronic systems into embedded computers

• MODULE FOR ADJUSTING PARAMETERS OF ALGORITHMS FOR AUTOMATIC DETECTION AND TRACKING OF OBJECTS FOR OPTOELECTRONIC SYSTEMS

  V. А. Tupikov, V. А. Pavlova, А.I. Lizin, P.А. Gessen

  71-81

  2022-04-20

  Abstract ▼

  In order to create an innovative module for automatic correction of algorithms for automatic
  detection and tracking of objects with real-time training, a study of world experience in the field
  of general-purpose automatic tracking with the ability to recognize the tracking object for use in
  embedded computing devices of optoelectronic systems of promising robotic complexes was carried
  out. Based on the conducted research, methods and approaches have been selected and tested
  that allow with the greatest accuracy, while maintaining high computational efficiency, to provide
  on-the-fly training of classifiers (online learning) without a priori knowledge of the type of tracking object and to ensure subsequent correction during tracking and detection of the original object
  in case of its short-term loss. Such methods include a histogram of directional gradients – a descriptor
  of key features based on the analysis of the distribution of brightness gradients of an object
  image. Its use allows you to reduce the amount of information used without losing key data
  about the object and increase the speed of image processing. The article substantiates the choice
  of one of the classification algorithms in real time, which allows solving the problem of binary
  classification - the method of support vectors. Due to the high speed of data processing and the
  need for a small amount of initial training data to build a separating hyperplane, on the basis of
  which the classification of objects takes place, this method is chosen as the most suitable for solving
  the task. To implement online training, a modification of the support vector machine was chosen,
  implementing stochastic gradient descent at each step of the algorithm – Pegasos. Another
  auxiliary method is the clustering method of key points – this ensures an accelerated selection of
  objects for classification and training. The authors of the study carried out the development and
  semi-natural modeling of the proposed module, evaluated the effectiveness of its work in the tasks
  of correcting and detecting the object of interest in real time with preliminary online training in
  the process of tracking the object. The developed algorithm has shown high efficiency in solving
  the problem. In conclusion, proposals are presented to further improve the accuracy and probability
  of detecting an object of interest by the developed algorithm, as well as to improve its performance
  by optimizing calculations.