Video Data Research Articles

Концептуальная структура системы поддержки принятия решений ситуационной видеоаналитики для многообъектного распознавания

Information is provided on the growing interest in situational video analytics (SVA) technologies in various fields of activity. The paper describes a trend towards the implementation of SVA systems based on cognitive artificial intelligence technologies, which make it possible to detect an object and its location in a video stream in real time with a high degree of accuracy. The architecture of a decision support system (DSS) with the possibility of multi-object recognition is proposed. The key logical components of the DSS in the field of SVA are highlighted, as well as the functions and purpose of each component are described. The role of the decision maker for implementing multi-object recognition when searching for an object in a video stream is particularly emphasized. The classification of neural networks by types and fields of application is given, and it is revealed that convolutional neural networks are used to solve video analytics problems. Examples of using the YOLOv5 convolutional model in the DSS model database management system in IAS tasks for detecting the presence of objects on video data are considered. A block diagram of the SVA object recognition algorithm has been developed. A series of experiments has been conducted to train a convolutional neural network model on a unique dataset, on an expanded dataset, and with a new additional object. This organization of experiments is aimed at improving the quality and accuracy of object recognition and exploring the possibility of multi-object recognition. As a result of the experiment, the final trained neural network model was tested and its potential capabilities were analyzed for use in the SVA DSS, taking into account the accuracy of the model. The accuracy was 91.6% for the validation set containing 2 objects. The results obtained using a trained neural network of the YOLOv5 architecture confirms the importance of convolutional neural networks as a key component of the database of SVA DSS models.

Evaluating the safety impact of mid-block pedestrian signals (MPS).

The Florida Department of Transportation (FDOT) has recently started implementing a new signal system at mid-blocks called Mid-block Pedestrian Signals (MPS). This study aims to evaluate the effectiveness of these newly implemented MPSs. A total of 260h of video data were collected from five locations across Florida, with 130h recorded before MPS installation and 130h after installation, including both weekdays and weekends. State-of-the-art computer vision technology was employed to detect and track various road users. A random parameters multinomial logit model with heterogeneity in the means was implemented to assess safety of vehicle-pedestrian interaction by three conflict categories: No Conflict, Moderate Conflict, and Serious Conflict. Relative-Time-to-Collision (RTTC) values were utilized to classify these level of conflicts. The analysis demonstrates that the presence of MPS significantly enhances safety outcomes by increasing the likelihood of avoiding conflicts and reducing the probabilities of both moderate and serious conflicts. Key factors influencing conflict probabilities were identified, including pedestrian and vehicle counts, average leading vehicle speed, standard deviation of leading vehicle speeds, and land-use mix, all of which increase the probability of serious conflicts. Interestingly, the analysis identified three significant interaction variables with MPS: average leading vehicle speed, standard deviation of leading vehicle speeds, and land-use mix. While these factors individually had a higher probability of leading to serious conflicts, the presence of MPS effectively mitigates these risks by moderating their adverse effects, increasing the likelihood of no conflicts. These results underscore the importance of MPS as an effective measure to improve safety at mid-block crossings.

Innovative Video Classification Method based on Deep Learning Approach

Background: The method includes: receiving a set of video data and labeling it into categories, segmenting the received videos into N segments, randomly selecting M frames for each video segment in the training phase, concatenating the video images into multi-channel images, and rolling. Methods: This work was developed in the Python programming language using the Keras library with Tensorflow as the back-end. The objective is to develop a network that presents performance compatible with the state of the art in terms of classifying videos according to the actions taken. Results: Given the hardware limitations, there is considerable distance between the implementation possibilities in this work and what is known as the state-of-the-art. Conclusion: Throughout the work, some aspects in which this limitation influenced the development are presented, but it is shown that this realization is feasible and that obtaining expressive results is possible. 98.6% accuracy is obtained in the UCF101 data set, compared to the 98 percentage points of the best result ever reported, using, however, considerably fewer resources. In addition, the importance of transfer learning in achieving expressive results as well as the different performances of each architecture are reviewed. Thus, this work may open doors to carry patent- based outcomes.

A comparative analysis of LSTM models aided with attention and squeeze and excitation blocks for activity recognition.

Human Activity Recognition plays a vital role in various fields, such as healthcare and smart environments. Traditional HAR methods rely on sensor or video data, but sensor-based systems have gained popularity due to their non-intrusive nature. Current challenges in HAR systems include variability in sensor data influenced by factors like sensor placement, user differences, and environmental conditions. Additionally, imbalanced datasets and computational complexity hinder the performance of these systems in real-world applications. To address these challenges, this paper proposes an LSTM-based HAR model enhanced with attention and squeeze-and-excitation blocks. The LSTM captures temporal dependencies, while the attention mechanism dynamically focuses on important parts of the input sequence. The squeeze-and-excitation block recalibrates channel-wise feature importance, allowing the model to emphasize the most informative features. The proposed model demonstrated a 99% accuracy rate, showcasing its effectiveness in recognizing various activities from sensor data. The integration of attention and squeeze-and-excitation mechanisms further boosted the model's ability to handle complex datasets. Comparative analysis with existing LSTM models confirms that the proposed approach improves accuracy and reduces computational complexity, making it a highly suitable model for real-world applications.

3세 유아반에서의 돌봄의 의미에 대한 포토에세이 탐구: 포스트휴머니즘의 관점에서

This study aimed to examine the stories of caring that are being expressed in the early childhood educational institutions where young children and teachers live together, and explored the phenomenon of caring and its meaning through photo essay exploration. To this end, the three-year-old class at the P city daycare center was observed 14 times and photographs and video data, research notes, and recorded conversations with teachers were collected. The photo essays for interpreting the meaning of the data were written through diffractive writing by re-viewing the images created through cutting. And the written photo essays went through a process of being reconstructed through re-reading and together-reading. The results of the study showed that the caring of three-year-old class from the perspective of posthumanism was ‘the entanglement of caring that crosses the boundaries of dichotomy’, ‘the transformation as a caregiver-becoming along with the agency of matter’, and ‘caring extended to non-human beings’. The results of this study suggest that acts of caring occurs when the act of care crosses within the phenomenon, that materials actively participate in the field of caring with agency, and that beings beyond humans and young children are creating caring together through relationships.

Generating context-specific sports training plans by combining generative adversarial networks.

Personalized sports training plans are essential for addressing individual athlete needs, but traditional methods often need to integrate diverse data types, limiting adaptability and effectiveness. Existing machine learning (ML) and rule-based approaches cannot dynamically generate context-specific training programs, reducing their applicability in real-world scenarios. This study aims to develop a Generative Adversarial Network (GAN)- based framework to create context-specific training plans by integrating numeric attributes (e.g., age, heart rate) and motion features from video data. The research focuses on improving context-specific efficiency and real-time adaptability while addressing the limitations of traditional methods. The proposed GAN framework combines numeric and motion features using a generator-discriminator architecture to produce tailored training plans. The model is evaluated quantitatively through metrics like mean square error (MSE) and generation time and qualitatively through subjective ratings from athletes and coaches using a five-point Likert scale for context-specific, scientificity, applicability, and feasibility. Statistical significance is analyzed using ANOVA testing. The proposed GAN model outperforms traditional ML and rule-based methods, achieving a 22% reduction in MSE and a 45% improvement in generation time. Subjective evaluations reveal significant improvements in context-specific and applicability, with ratings averaging 4.8/5 compared to 3.9/5 for baseline models. The GAN framework effectively integrates multimodal data, demonstrating dynamic adaptability and high efficiency suitable for real-world applications. The proposed GAN-based framework advances the generation of personalized sports training plans by integrating numeric and motion data, achieving superior adaptability and efficiency. These results highlight the model's potential for practical deployment in athletic coaching systems, addressing critical gaps in existing methodologies and offering scalable solutions for individualized training.

Experimental implementation of chaos-based video encryption using visible light communication technology

Simulations are commonly used to develop and evaluate video encryption algorithms. Although these approaches are useful for demonstrating theoretical feasibility and algorithm performance, they neglect practical challenges and real-world communication conditions. This oversight creates a critical gap in evaluating their effectiveness and security in practical applications. In this paper, we describe a novel, to the best of our knowledge, video encryption scheme that employs a modified chaotic Colpitts oscillator, a customized Josephus problem, and ribonucleic acid (RNA) operations, and is implemented in visible light communication (VLC) environment. The modified Colpitts oscillator reveals significant phenomena, the most prominent for this study being chaos. The encryption scheme involves block scrambling and pixel-level permutation employing a customized Josephus problem and a column and row shifting technique. Diffusion is then achieved through a custom Josephus problem-oriented RNA operation combined with basic RNA operations and cipher block chaining (CBC). Based on the results, we found that the customized Josephus problem-based permutation performed more efficiently than traditional methods, whereas the column and row shifting permutation was more robust. By eliminating key sequence coding and item-to-item operations, the proposed RNA operation reduces computational overhead. This video encryption scheme is proven to be effective and resilient at securing video data transmitted over the VLC channel, and it constitutes a significant advancement in areas such as surveillance, medical imaging, and military communication requiring robust secured video data.

Influence of the Camera Viewing Angle on OpenPose Validity in Motion Analysis

(1) Background: With human pose estimation on the rise in the field of biomechanics, the need for scientific investigation of those algorithms is becoming evident. The validity of several of those algorithms has been presented in the literature. However, there is only limited research investigating the applicability of human pose estimation outside the lab. The aim of this research was to quantify the effect of deviating from the standard camera setup used in biomechanics research. (2) Methods: Video data from four camera viewing angles were recorded and keypoints estimated using OpenPose. Kinematic data were compared against a gold-standard marker-based motion capture system to quantify the effect of the camera viewing angle on the validity of joint angle estimation of the knee, hip, elbow and shoulder joints. (3) Results: The results of this study showed reasonable correlations between the joint angles of OpenPose and the gold standard, except for the shoulder. However, the analysis also revealed significant biases when comparing the joint angles inferred from the different viewing angles. In general, back-viewing cameras performed best and resulted in the lowest percental deviations. (4) Conclusions: The findings of this study underscore the importance of conducting a detailed examination of individual movements before proposing specific camera angles for users in diverse settings.

Prospects for the Use of Machine Learning for Mood Disorders

Relevance. Mental disorders are one of the key medical and social issues. Over the last years artificial intelligence (AI) methods including machine and deep learning have been actively developing. This narrative review aimed to identify current and promising areas for the development and application of AI into clinical practice using the example of patients with depression and bipolar disorder. Methods. The search of publications was performed in January ─ February 2024 in PubMed, Google Scholar, elibrary databases using the combination of keywords: psychiatry, mental health, psychiatric disorder, depression, depressive episode, major depressive disorder, bipolar disorder, machine learning, deep learning, artificial intelligence. The review included original articles on the use of AI methods in patients with depression and bipolar disorder, and review articles devoted to the problems of applying AI methods in psychiatry and published in Russian or English in the last 10 years. Results. Most often, neuroimaging (mainly MRI and EEG), text, audio and video data, electronic device data, molecular genetics, clinical data and its combination, are used for machine learning (ML) models in patients with mood disorders. Despite the potential benefits of AI application in psychiatry, its implementation into clinical practice is currently challenging due to number of difficulties, such as small sample sizes, low representativeness, lack of standardization, inclusion of “noise” and correlated variables into models, lack of model testing on independent samples. Conclusion. Studies of ML methods have shown promising results for the early diagnosis of affective episodes and predicting response to therapy. However, the use of ML methods into clinical practice has a number of limitations, primarily due to insufficient validation. There is a need for well-designed prospective cohort studies, as well as the development of extensive and high-quality databases and models capable of identifying new relationships between variables in order to overcome these limitations.

Diagnosis of depression based on facial multimodal data

IntroductionDepression is a serious mental health disease. Traditional scale-based depression diagnosis methods often have problems of strong subjectivity and high misdiagnosis rate, so it is particularly important to develop automatic diagnostic tools based on objective indicators.MethodsThis study proposes a deep learning method that fuses multimodal data to automatically diagnose depression using facial video and audio data. We use spatiotemporal attention module to enhance the extraction of visual features and combine the Graph Convolutional Network (GCN) and the Long and Short Term Memory (LSTM) to analyze the audio features. Through the multi-modal feature fusion, the model can effectively capture different feature patterns related to depression.ResultsWe conduct extensive experiments on the publicly available clinical dataset, the Extended Distress Analysis Interview Corpus (E-DAIC). The experimental results show that we achieve robust accuracy on the E-DAIC dataset, with a Mean Absolute Error (MAE) of 3.51 in estimating PHQ-8 scores from recorded interviews.DiscussionCompared with existing methods, our model shows excellent performance in multi-modal information fusion, which is suitable for early evaluation of depression.

FL-TENB4: A Federated-Learning-Enhanced Tiny EfficientNetB4-Lite Approach for Deepfake Detection in CCTV Environments

The widespread deployment of CCTV systems has significantly enhanced surveillance and public safety across various environments. However, the emergence of deepfake technology poses serious challenges by enabling malicious manipulation of video footage, compromising the reliability of CCTV systems for evidence collection and privacy protection. Existing deepfake detection solutions often suffer from high computational overhead and are unsuitable for real-time deployment on resource-constrained CCTV cameras. This paper proposes FL-TENB4, a Federated-Learning-enhanced Tiny EfficientNetB4-Lite framework for deepfake detection in CCTV environments. The proposed architecture integrates Tiny Machine Learning (TinyML) techniques with EfficientNetB4-Lite, a lightweight convolutional neural network optimized for edge devices, and employs a Federated Learning (FL) approach for collaborative model updates. The TinyML-based local model ensures real-time deepfake detection with minimal latency, while FL enables privacy-preserving training by aggregating model updates without transferring sensitive video data to centralized servers. The effectiveness of the proposed system is validated using the FaceForensics++ dataset under resource-constrained conditions. Experimental results demonstrate that FL-TENB4 achieves high detection accuracy, reduced model size, and low inference latency, making it highly suitable for real-world CCTV environments.

Listening to the Formation of Polymorphs in a Ball Mill.

In this contribution, we apply our newly developed ball-milling operando platform, which combines Raman spectroscopy and thermal (IR) imaging, as well as acoustic and high-speed optical video recordings, to the synthesis and transformation of citric acid-isonicotinamide (1:2) cocrystal polymorphs in transparent PMMA jars. Particularly, we demonstrate how Raman, temperature, acoustic, and video data are complementary and enable detection and connection of chemical and physical events happening during ball-milling in a time-resolved manner. Importantly, we show that the formation of the three cocrystal polymorphs can be detected through acoustic analyses solely. Even more impressively, the meticulous analyses of the sound data allowed subtle polymorphic transitions to be perceived in operando mode when the Raman spectroscopy was not conclusive enough. Such changes in sound signatures are greatly linked to the beads' motions in the milling jar. A new data analysis methodology of acoustic recordings is proposed through a combination of energetic and statistical approaches that simplifies the data analyses for potential users. The interpretation of the detected sound signals was further validated thanks to the high-speed videos recorded in synchronization with all other operando techniques. Finally, we broaden this acoustic methodology to opaque stainless-steel jars, showing the relevance of the acoustic analysis method for following polymorphic transformations of cocrystals, as well as pure substances, in any type of milling jar.

The Relationship between Classroom Learning Engagement of Secondary Vocational School Students and Psychological Factors based on Video Analysis Technology

In the quest for quality improvement in higher education, student engagement in the classroom as a core element in evaluating the effectiveness of education has become increasingly important. In particular, the precise measurement and analysis of students' engagement in learning behaviors in the classroom have become a key area for combining learning engagement research with teaching practice in higher education institutions. This study focuses on the use of computer video parsing technology to detect the learning engagement of students in secondary vocational schools in Jiangsu Province (secondary students) and its application. The study analyses the video data generated by secondary students in the learning process, and extracts key information with the help of advanced computer video parsing algorithms, to accurately quantify students' learning engagement. The results of the study show that this technology can effectively monitor the learning engagement of secondary vocational school students, and provide an effective aid for teaching improvement and personalized learning guidance.

Full fine-tuning strategy for endoscopic foundation models with expanded learnable offset parameters.

In the medical field, endoscopic video analysis is crucial for disease
diagnosis and minimally invasive surgery. The Endoscopic Foundation Models (Endo-
FM) utilize large-scale self-supervised pre-training on endoscopic video data and
leverage video transformer models to capture long-range spatiotemporal dependencies.
However, detecting complex lesions such as gastrointestinal metaplasia (GIM) in
endoscopic videos remains challenging due to unclear boundaries and indistinct
features, and Endo-FM has not demonstrated good performance. To this end, we
propose a fully fine-tuning strategy with an Extended Learnable Offset Parameter
(ELOP), which improves model performance by introducing learnable offset parameters
in the input space. Specifically, we propose a novel loss function that combines cross-
entropy loss and focal loss through a weighted sum, enabling the model to better focus
on hard-to-classify samples during training. We validated ELOP on a private GIM
dataset from a local grade-A tertiary hospital and a public polyp detection dataset.
Experimental results show that ELOP significantly improves the detection accuracy,
achieving accuracy improvements of 6.25 % and 3.75%respectively compared to the
original Endo-FM. In summary, ELOP provides an excellent solution for detecting
complex lesions in endoscopic videos, achieving more precise diagnoses.

A novel approach to frame reduction in robust hand gesture recognition

Abstract Hand gesture recognition is becoming an increasingly integral part of our daily lives, enabling seamless communication, enhancing interaction, and revolutionizing multiple industries. To ensure a more precise and efficient system, the key aspect of hand gestures lies in detecting hand patterns and retrieving the hand gestures. However, as the volume of video data increases, extracting the essential hand patterns while excluding unnecessary frames becomes a challenge. Addressing this issue, a novel Harris Hawk Optimization K-Means frame reduction is proposed, inspired by the hunting behavior of Harris Hawks in nature. This proposed approach combines the Harris Hawk Optimization algorithm with the K-Means clustering method. The algorithm simulates the hunting behavior of Harris Hawks and utilizes Euclidean distance as a fitness function to determine the optimal frames. Subsequently, the K-Means clustering method is employed to group similar frames together based on these optimal selections. An average frame is generated and aggregated for each cluster to form a reduced set of frames. These reduced frames are then classified using the modified Mobilenet V2 model, outperforming other state-of-the-art techniques by achieving an exceptional accuracy rate of 99.93%. The experiment results lay the groundwork for incorporating the novel framework of hand gesture recognition into a range of applications, including sign language interpretation, human-computer interaction, and virtual reality systems.

Edge-Based Dynamic Spatiotemporal Data Fusion on Smart Buoys for Intelligent Surveillance of Inland Waterways

Increasing vessel traffic in narrow, winding inland waterways has heightened the risk of accidents, driving the need for improved surveillance and management. This study addresses the challenge of real-time processing and synchronization of voluminous video and AIS data for effective waterway management. We developed a surveillance method utilizing smart buoys equipped with sensors and edge computing devices, enabling dynamic spatiotemporal data fusion. The integration of AIS data with advanced computer vision techniques for target detection allows for real-time traffic analysis and provides detailed navigational dynamics of vessels. The method employs an enhanced Long Short-Term Memory network for precise trajectory prediction of AIS data and a single-stage target detection model for video data analysis. Experimental results demonstrate significant improvements in ship detection accuracy and tracking precision, with an average position prediction error of approximately 1.5 m, which outperforms existing methods. Additionally, a novel regional division and a Kalman filter-based method for AIS and video data fusion were proposed, effectively resolving the issues of data sparsity and coordinate transformation robustness under complex waterway conditions. This approach substantially advances the precision and efficiency of waterway monitoring systems, providing a robust theoretical and practical framework for the intelligent supervision of inland waterways.

Brain-inspired multimodal motion and fine-grained action recognition

IntroductionTraditional action recognition methods predominantly rely on a single modality, such as vision or motion, which presents significant limitations when dealing with fine-grained action recognition. These methods struggle particularly with video data containing complex combinations of actions and subtle motion variations.MethodsTypically, they depend on handcrafted feature extractors or simple convolutional neural network (CNN) architectures, which makes effective multimodal fusion challenging. This study introduces a novel architecture called FGM-CLIP (Fine-Grained Motion CLIP) to enhance fine-grained action recognition. FGM-CLIP leverages the powerful capabilities of Contrastive Language-Image Pretraining (CLIP), integrating a fine-grained motion encoder and a multimodal fusion layer to achieve precise end-to-end action recognition. By jointly optimizing visual and motion features, the model captures subtle action variations, resulting in higher classification accuracy in complex video data.Results and discussionExperimental results demonstrate that FGM-CLIP significantly outperforms existing methods on multiple fine-grained action recognition datasets. Its multimodal fusion strategy notably improves the model's robustness and accuracy, particularly for videos with intricate action patterns.

The Analysis of Emotion-Aware Personalized Recommendations via Multimodal Data Fusion in the Field of Art

This paper proposes an emotion-aware personalized recommendation system (EPR-IoT) based on IoT data and multimodal emotion fusion, aiming to address the limitations of traditional recommendation systems in capturing users' emotional states of artistic product consumption in real time. With the proliferation of smart devices, physiological signals such as heart rate and skin conductance—which are strongly correlated with emotional states—provide new opportunities for emotion recognition. For example, an increase in heart rate is typically associated with emotions like anxiety, anger, or fear, while a decrease is linked to emotional states like relaxation or joy. Similarly, skin conductance rises with emotional arousal, particularly during stress or fear. These physiological signals, combined with text, speech, and video data of art products, are fused to construct an art emotion-driven recommendation model capable of dynamically adjusting the recommended content.

P0567 Use of a convolutional neural network to quantify and predict ulcerative colitis disease endoscopic activity and histological outcomes

Abstract Background There is considerable inter-user variability in endoscopic and histologic scoring of the severity of ulcerative colitis (UC). The aim of this study was to create and train an artificial intelligence algorithm to predict the severity of UC using endoscopic videos. Methods This prospective cohort study recruited UC patients undergoing flexible sigmoidoscopy or colonoscopy between October 2023 and September 2024. The endoscopic examination was recorded from the splenic flexure to the anus. The videos were segmented into 3 second tubelets, and poor quality videos excluded. Objective disease activity was quantified by an experienced IBD gastroenterologist (by determining individual endoscopic Mayo Scores- eMS), and corresponding biopsies were scored according to the Nancy Histologic Index (NHI) by one of three gastrointestinal histopathologists. The dataset was divided into a training dataset (38 patients, 88% of video data) and a test dataset (6 patients, 12% of data), stratified by disease severity and equipment source. Two video vision transformer models were trained to recognise Mayo and NHI respectively, and their performance in differentiating disease activity from remission, and accuracy in determining endoscopic and histologic severity scores were evaluated. Standard diagnostic accuracy measures were used to calculate the models’ ability to detect and quantitate UC disease activity and included area under the receiver operating characteristic curve (AUROC), binary and multi-class accuracy, sensitivity, specificity, and precision. Results 53 patients were recruited, and after exclusions a total of 44 were included in the final analysis. 30 patients had endoscopic activity and 30 patients had histological activity. The resultant eMS and NHI showed moderate correlation (R=0.72). For eMS, the accuracy of the model was excellent and had an AUROC of 0.98 in identifying disease activity (eMS≥1), with accuracy, sensitivity, specificity, and precision of 92%, 86%, 96%, 93% respectively. Multiclass accuracy was 87%, with a macro F1 score of 0.85. For NHI, the model was highly accurate in detecting histopathologic activity (NHI≥1), with AUROC of 0.95, accuracy, sensitivity, specificity, precision at 89%, 92%, 86%, 85% respectively. However, it was less robust in predicting specific NHI scores, with multiclass accuracy at 69%, and a macro F1 score of 0.41. Conclusion We developed an AI algorithm that is able to reliably predict endoscopic and histologic disease activity for patients with UC, and accurately differentiate those patients from those in remission. This may reduce clinician reliance on biopsies to determine disease activity, and have far-reaching implications in UC disease assessment. References Hashash JG, Yu Ci Ng F, Farraye FA, Wang Y, Colucci DR, Baxi S, Muneer S, Reddan M, Shingru P, Melmed GY. Inter- and Intraobserver Variability on Endoscopic Scoring Systems in Crohn's Disease and Ulcerative Colitis: A Systematic Review and Meta-Analysis. Inflamm Bowel Dis. 2024 Mar 28:izae051. doi: 10.1093/ibd/izae051 Marchal-Bressenot A, Salleron J, Boulagnon-Rombi C, et al. Development and validation of the Nancy histological index for UC. Gut 2017;66:43-49.

Optimizing IoT Video Data: Dimensionality Reduction for Efficient Deep Learning on Edge Computing

The rapid loss of biodiversity significantly impacts birds’ environments and behaviors, highlighting the importance of analyzing bird behavior for ecological insights. With the growing adoption of Machine Learning (ML) algorithms in the Internet of Things (IoT) domain, edge computing has become essential to ensure data privacy and enable real-time predictions by processing high-dimensional data, such as video streams, efficiently. This paper introduces a set of dimensionality reduction techniques tailored for video sequences based on cutting-edge methods for this data representation. These methods drastically compress video data, reducing bandwidth and storage requirements while enabling the creation of compact ML models with faster inference speeds. Comprehensive experiments on bird behavior classification in rural environments demonstrate the effectiveness of the proposed techniques. The experiments incorporate state-of-the-art deep learning techniques, including pre-trained video vision models, Autoencoders, and single-frame feature extraction. These methods demonstrated superior performance to the baseline, achieving up to a 6000-fold reduction in data size while reaching a classification accuracy of 60.7% on the Visual WetlandBirds Dataset and obtaining state-of-the-art performance on this dataset. These findings underline the potential of using dimensionality reduction to enhance the scalability and efficiency of bird behavior analysis.