10 research outputs found
Training Sample Formation for Convolution Neural Networks to Person Re-Identification from Video
No full textTo improve the person re-identification system accuracy, an integrated approach is proposed in the formation of a training sample for convolutional neural networks, which involves the use of a new image dataset, an increase in the training examples number using existing datasets, and the use of a number of transformations to increase their diversity. The created dataset PolReID1077 contains images of people that were obtained in all seasons, which will improve the correct operation of re-identification systems when the seasons change. Another PolReID1077 advantage is the video data use obtained from external and internal surveillance in a large number of different filming locations. Therefore, the people images in the created set are characterized by the variability of the background, brightness and color characteristics. Joining the created dataset with the existing CUHK02, CUHK03, Market-1501, DukeMTMC-ReID and MSMT17 sets made it possible to obtain 109 772 images for training. An increase in the variety of generated examples is achieved by applying a cyclic shift to them, eliminating color and replacing a fragment with a reduced copy of another image. The research results on estimating the accuracy of re-identification for the ResNet-50 and DenseNet-121 convolutional neural networks during their training, using the proposed approach to form a training sample, are presented
Organization Principles and Approaches Analysis to Improving the Person Re-Identification Accuracy in Distributed Video Surveillance Systems
No full textПриведена классификация существующих систем повторной идентификации по таким критериям, как тип системы, количество и вид запросов, время работы. Рассмотрена общая схема, отражающая основной принцип работы систем повторной идентификации, а также основные подходы и методы для решения этой задачи с использованием сверточных нейронных сетей. Выполнено исследование существующих способов повышения точности работы алгоритмов и систем повторной идентификации. Проведен анализ влияния выбора гиперпараметров при обучении сверточных нейронных сетей на эффективность и динамику обучения алгоритма повторной идентификации.The paper presents a classification of existing re-identification systems according to such criteria as system type, requests number and type, and operating time. The general scheme is discussed, which reflects the basic operation principle of re-identification systems, and the main approaches and methods for solving this problem using convolutional neural networks are considered. The study ways existing to improve re-identification algorithms and systems accuracy has been carried out. The influence analysis hyperparameters choice in convolutional neural networks training on the efficiency and dynamics re-identification algorithm training is carried out
Person re-identification accuracy improvement by training a CNN with the new large joint dataset and re-rank
No full textThe paper is aimed to improve person re-identification accuracy in distributed video surveillance systems based on constructing a large joint image dataset of people for training convolutional neural networks (CNN). For this aim, an analysis of existing datasets is provided. Then, a new large joint dataset for person re-identification task is constructed that includes the existing public datasets CUHK02, CUHK03, Market, Duke, MSMT17 and PolReID. Testing for re-identification is performed for such frequently cited CNNs as ResNet-50, DenseNet121 and PCB. Re-identification accuracy is evaluated by using the main metrics Rank, mAP and mINP. The use of the new large joint dataset makes it possible to improve Rank1 mAP, mINP on all test sets. Re-ranking is used to further increase the re-identification accuracy. Presented results confirm the effectiveness of the proposed approach
Joint Dataset for CNN-based Person Re-identification
No full textIn this paper, we propose a joint dataset for person re-identification task that includes the existing public datasets CUHK02, CUHK03, Market, Duke, LPW and our collected PolReID. We investigate the training dataset size and composition effect on the re-identification accuracy. We carried out a number of experiments with different size of dataset to solve re-identification task. The results of experiments are presented
On Whitney-type Characterization of Approximate Differentiability on Metric Measure Spaces
No full textWe study approximately differentiable functions on metric measure spaces admitting a Cheeger differentiable structure. The main result is a Whitney-type characterization of approximately differentiable functions in this setting. As an application, we prove a Stepanov-type theorem and consider approximate differentiability of Sobolev, BV and maximal functions.Peer reviewe
Формирование обучающей выборки для свёрточных нейронных сетей при реидентификации людей по видеоданным
No full textTo improve the person re-identification system accuracy, an integrated approach is proposed in the formation of a training sample for convolutional neural networks, which involves the use of a new image dataset, an increase in the training examples number using existing datasets, and the use of a number of transformations to increase their diversity. The created dataset PolReID1077 contains images of people that were obtained in all seasons, which will improve the correct operation of re-identification systems when the seasons change. Another PolReID1077 advantage is the video data use obtained from external and internal surveillance in a large number of different filming locations. Therefore, the people images in the created set are characterized by the variability of the background, brightness and color characteristics. Joining the created dataset with the existing CUHK02, CUHK03, Market-1501, DukeMTMC-ReID and MSMT17 sets made it possible to obtain 109 772 images for training. An increase in the variety of generated examples is achieved by applying a cyclic shift to them, eliminating color and replacing a fragment with a reduced copy of another image. The research results on estimating the accuracy of re-identification for the ResNet-50 and DenseNet-121 convolutional neural networks during their training, using the proposed approach to form a training sample, are presented.Для повышения точности работы системы реидентификации людей предлагается комплексный подход при формировании обучающей выборки для свёрточных нейронных сетей, предполагающий использование нового набора изображений, увеличение количества тренировочных примеров за счет существующих баз данных и применение ряда преобразований для повышения их разнообразия. Созданный набор данных PolReID1077 содержит изображения людей, которые были получены во все времена года, что позволит повысить корректность работы систем реидентификации при смене сезонов. ПреимуществомPolReID1077 является также использование видеоданных, полученных при внешнем и внутреннем наблюдении в большом количестве различных мест съемки. Поэтому изображения людей в созданном наборе характеризуются вариабельностью фона, яркостных и цветовых характеристик. Объединение созданного набора с существующими CUHK02, CUHK03, Market-1501, DukeMTMC-ReID и MSMT17 позволило получить 109 772 изображения для обучения. Увеличение разнообразия сформированных примеров достигается за счет применения к ним циклического сдвига, исключения цветности и замещения фрагмента уменьшенной копией другого изображения. Представлены результаты исследований по оценке точности реидентификации для свёрточных нейронных сетей ResNet-50 и DenseNet-121 при их тренировке с использованием предложенного подхода для формирования обучающей выборки
Person Re-Identificaion Algorithm by Image from Video Surveilance Sistem Using a Neural Network Compound Descriptor
No full textPerson re-identification in video surveillance systems by feature replacement of occluded parts of human figures
No full textТhe algorithm for re-identifying people in intelligent video surveillance systems is proposed. It is based on the construction of a compound neural network descriptor and replacing features of occluded parts of a human figure. Composite descriptors are generated for all images in a gallery and recorded in a table. They characterize the global and local features of each person, considering his individual parts’ visibility. Detection and selection of areas of interest for the formation of local descriptors is carried out based on detecting key points of the human body. If a person is partially occluded by other people or objects, then the corresponding region is classified as invisible, and the compound descriptor of respective component will be invalid and equal to zero. For images whose feature vector has zero components, the feature table is ranked according to the cosine similarity metric for each visible local fragment. Based on the feature table rankings, the k-nearest neighbors are determined and the k1-best are selected from them. The corresponding k1-nearest neighbors’ component average value of the feature vector is used to replace the zero descriptor components. The feature table is then updated for the generated vectors and ranking is performed according to the query using the cosine similarity metric. ResNet-50 and DenseNet-121 were used as backbone CNNs for feature extraction, and testing was performed using Market-1501, DukeMTMC-ReID, Occluded-Duke, MSMT17, and PolReID1077 datasets
Person Re-identification in Video Surveillance Systems Using Deep Learning: Analysis of the Existing Methods
No full textThis paper is devoted to a multifaceted analysis of person re-identification (ReID) in video surveillance systems and modern solution methods using deep learning. The general principles and application of convolutional neural networks for this problem are considered. A classification of person ReID systems is proposed. The existing datasets for training deep neural architectures are studied and approaches to increasing the number of images in databases are described. Approaches to forming human image features are considered. The backbone models of convolutional neural network architectures used for person ReID are analyzed and their modifications as well as training methods are presented. The effectiveness of person ReID is examined on different datasets. Finally, the effectiveness of the existing approaches is estimated in different metrics and the corresponding results are given
Увеличение точности реидентификации людей на основе двухэтапного обучения сверточных нейронных сетей и аугментации
No full textObjectives. The main goal is to improve person re-identification accuracy in distributed video surveillance systems.Methods. Machine learning methods are applied.Result. A technology for two-stage training of convolutional neural networks (CNN) is presented, characterized by the use of image augmentation for the preliminary stage and fine tuning of weight coefficients based on the original images set for training. At the first stage, training is carried out on augmented data, at the second stage, fine tuning of the CNN is performed on the original images, which allows minimizing the losses and increasing model efficiency. The use of different data at different training stages does not allow the CNN to remember training examples, thereby preventing overfitting.Proposed method as expanding the training sample differs as it combines an image pixels cyclic shift, color exclusion and fragment replacement with a reduced copy of another image. This augmentation method allows to get a wide variety of training data, which increases the CNN robustness to occlusions, illumination, low image resolution, dependence on the location of features.Conclusion. The use of two-stage learning technology and the proposed data augmentation method made it possible to increase the person re-identification accuracy for different CNNs and datasets: in the Rank1 metric by 4–21 %; in the mAP by 10–31 %; in the mINP by 39–60 %. Цели. Основной целью является повышение точности повторной идентификации людей в распределенных системах видеонаблюдения.Методы. Используются методы машинного обучения.Результаты. Представлена технология двухэтапного обучения сверточных нейронных сетей (СНС), отличающаяся использованием аугментации изображений для предварительного этапа и точной настройки весовых коэффициентов на основе исходного набора изображений. На первом этапе обучение осуществляется на аугментированных данных, затем выполняется точная настройка СНС на исходных изображениях, что способствует повышению эффективности ре-идентификации за счет уменьшения потерь при обучении. Использование на двух этапах разных данных не позволяет СНС запоминать тренировочные примеры, тем самым предотвращая переобучение.Предложенный метод расширения набора данных для обучения отличается тем, что совмещает циклический сдвиг пикселей изображения, исключение цветности и замещение фрагмента уменьшенной копией другого из пакета, подаваемого на вход СНС. Данный метод аугментации позволяет увеличить разнообразие обучающих данных, что повышает робастность СНС ко многим факторам: перекрытию людей, изменению освещенности, уменьшению разрешения изображения, зависимости от местоположения отличительных особенностей объекта интереса.Заключение. Применение технологии двухэтапного обучения и предложенного метода аугментации данных позволило повысить точность повторной идентификации людей для разных СНС и наборов данных в метриках: Rank1 на 4% – 21%; mAP на 10% – 31%; mINP на 39% – 60%
