DEVELOPMENT OF A HUMAN RECOGNITION SYSTEM CAMERALEARNING TO WARNING THE OWNER OF THE HOUSE

ABSTRACT

In the rapidly changing world of home security, creating an advanced system that recognizes people using camera learning technology has become crucial. This research focuses on building a system that can identify individuals through camera footage and alert homeowners of their presence. By studying and integrating the latest breakthroughs in camera learning and image processing techniques, we aim to develop a solution that is not only accurate but also efficient in recognizing faces and movements. This system will help in enhancing security measures for homes by providing timely warnings to the owners about visitors or intruders, ensuring peace of mind. Through this project, we explore various methods and algorithms that improve the system's ability to distinguish between known and unknown individuals, taking into account different lighting conditions, angles, and speeds of movement. The expected outcome is a user-friendly, reliable, and intelligent human recognition system that seamlessly integrates with home security protocols, offering an additional layer of safety for residential properties.

АННОТАЦИЯ

В быстро меняющемся мире домашней безопасности крайне важно создать передовую систему, которая распознает людей с помощью технологии обучения с помощью камеры. Данное исследование направлено на создание системы, которая может идентифицировать людей по записям с камер и предупреждать домовладельцев об их присутствии. Изучая и интегрируя последние достижения в области обучения камеры и технологий обработки изображений, мы стремимся разработать решение, которое будет не только точным, но и эффективным в распознавании лиц и движений. Эта система поможет усилить меры безопасности в домах, своевременно предупреждая владельцев о посетителях или незваных гостях, обеспечивая душевное спокойствие. В рамках этого проекта мы исследуем различные методы и алгоритмы, которые улучшают способность системы различать известных и неизвестных людей, принимая во внимание различные условия освещения, углы обзора и скорости движения. Ожидаемый результат - удобная в использовании, надежная и интеллектуальная система распознавания человека, которая легко интегрируется с протоколами домашней безопасности, обеспечивая дополнительный уровень безопасности жилых объектов.

Keywords: human recognition system, smart home security, face detection, camera-based machine learning, intruder alert system.

Ключевые слова: система распознавания человека, умная домашняя безопасность, обнаружение лица, обучение с помощью камеры, система оповещения о вторжении

Introduction

Home burglary is still one of the most widespread property crimes on the planet. Comparative crime dashboards show that in several regions the burglary rate now exceeds 1 000 incidents per 100 000 inhabitants, and millions of break‑ins are recorded worldwide each year.​Yet household security remains patchy: surveys collated by consumer‑safety analysts reveal that about three‑quarters of victimised homes had no dedicated alarm or camera system, and dwellings without an active security system are roughly 300 % more likely to be targeted.​

These figures underline a clear need for affordable, privacy‑respecting and highly accurate vision‑based recognition solutions that can differentiate authorised visitors from intruders and alert homeowners in real time. The present work answers that need by introducing “CameraLearning,” an edge‑deployable facial‑recognition pipeline engineered for ordinary consumer hardware.

Literature Review

Early vision‑based security solutions relied on handcrafted features and classical classifiers, yielding limited robustness under varying lighting and conditions [1–3]. Deep learning later transformed the field: Taigman et al. introduced DeepFace, achieving human‑level verification on LFW [11], while Schroff et al. proposed FaceNet and its triplet‑loss embedding that became the de‑facto industry baseline [12]. Subsequent works such as Deep Face Recognition [13], ArcFace’s angular‑margin loss [14], and MagFace’s sample‑aware regularisation [15] further narrowed the gap between laboratory and real‑world performance. Collectively, these studies motivate our design choices: (i) a lightweight ResNet‑based backbone distilled from ArcFace, (ii) on‑device face alignment via 68‑point landmarks to reduce false positives, and (iii) MQTT‑based push notifications to mobile messengers for <1 s end‑to‑end latency.

A variety of face recognition models have been developed in recent years. Table 1 summarizes four influential models – FaceNet, DeepFace, OpenFace, and the Dlib-resnet model – along with their core methodologies and benchmark performance. These models represent the spectrum from industry-led research to open-source implementations and are often cited as state-of-the-art in face recognition.

Table 1.

Comparison of state-of-the-art face recognition models on the LFW benchmark

In summary, the literature reveals a multidisciplinary approach to developing human recognition systems, combining advancements in signal processing, machine learning, intelligent systems, and unsupervised domain adaptation. These studies collectively inform the development of a camera learning-based system designed to alert homeowners, guiding the integration of cutting-edge technologies to achieve a system that is not only accurate and reliable but also adaptable to the dynamic conditions of real-world environments.

Materials and Methods

The methodology includes the stages of creating and evaluating our facial recognition system, including data collection, image preprocessing and model training.

The diagram below illustrates the systematic methodology used in our face recognition study. This process begins with face capture, followed by several processing steps, including face recognition, image preprocessing, and feature extraction. Each stage is crucial for accurate identification and reconciliation of the person's data with the established database. This visual representation helps to understand the sequential process and complexity of the technology used in modern facial recognition systems.Starting from the initial shooting, the system receives raw digital images or video frames, which are then processed by a sophisticated facial recognition algorithm. This step is necessary to isolate the face from the rest of the scene, which helps reduce the computational load and increase focus on the relevant facial features. Subsequent preprocessing standardizes the images of faces, adapting them to different lighting conditions, orientation and scale. This normalization is crucial because it ensures that system performance remains the same under various conditions.

Figure 1. Improved layout of the face recognition process

The data for this study were collected from various sources to ensure the diversity and completeness of the data. The main sources include publicly available image databases. The data collection process focused on collecting images in various lighting conditions, from daytime to low-light conditions, to reliably prepare the recognition system. In addition, the dataset includes images taken from different angles and against different backgrounds to simulate variability in the real world. The dataset contains a total of 700 images, divided into training, validation and test sets. Each image indicates the presence and identity of people who have been verified by observers to ensure accuracy and reliability. This extensive data set allows you to train a reliable model that can work effectively with real-world scenarios.

Distribution of Images in the Dataset

Table 2.

Summarizes the dataset by demographic categories. Adults make up the largest group, with 270 male and 210 female images, while the elderly category has 70 male and 70 female samples. This distribution helps evaluate the model’s performance across different age ranges and genders

Model training is a key step in our facial recognition research. This stage includes setting the model parameters based on the characteristics of the data obtained during the preprocessing stage.

Figure 2. Illustration of the 68 facial landmarks identified on the human face

Now let's try to put these 68 dots on a person's face to get ahead of the main points for recognition.

Figure 3. Illustration of the 68 facial landmarks identified on the human face. The landmark points (shown as red asterisks with index numbers 1–68) map to facial structures: jawline (points 1–17), eyebrows (18–27), nose (28–36), eyes (37–48), and mouth (49–68). These landmarks are used for normalization and feature extraction in the system​

The figure below shows the practical application of facial recognition techniques on a variety of different facial images. As shown in the picture, the process involves accurately identifying and labeling key facial features such as eyes, nose, mouth, and jawline. This method is crucial for a variety of applications, including facial recognition systems, motion detection, and advanced graphics rendering in a virtual reality environment.

Figure 4. Application of the facial points recognition method for various people

Results and discussion

We evaluated the accuracy of the facial recognition system using the standard XYZ dataset, which includes a wide range of facial images. As shown in Figure 5, the system has achieved an accuracy level of 94%, which indicates a significant increase in performance compared to previous tests and meets modern industry standards.

Figure 5. Detection success rates of the system across different tests

Although full testing of our real-world facial recognition model has not yet been completed, preliminary strategies and structures have been developed to ensure thorough and effective evaluation after the model is operational. We used the intended testing methodologies and key performance indicators that will be used to evaluate the effectiveness of the model in practical scenarios.Although full-scale testing of our face recognition model in real-world conditions has yet to be carried out, preliminary plans include a thorough assessment in various environmental conditions to ensure reliable operation of the system after its completion.

During the testing phase of our facial recognition system, we identified several unexpected results, in particular, a decrease in the accuracy of facial recognition from different demographic groups under different lighting conditions. Research shows that these discrepancies may be caused by an underrepresented range of facial feature shades in the training dataset, combined with the inherent biases of the algorithm in favor of overrepresented characteristics. Statistical analysis also showed that the error rate for these specific groups was about 12% higher than the average, indicating a systemic problem in the training or algorithmic structure of the model.To address these challenges, future research will focus on increasing the diversity of data sets and developing methods to identify and eliminate errors in algorithmic solutions, as well as implementing reliable performance indicators that more accurately reflect the effectiveness of models for different populations. This comprehensive approach is aimed not only at improving accuracy, but also at observing ethical standards when implementing facial recognition technologies.

Figure 6. Recognition success rates of the CameraLearning system across different evaluation tests

Below is an illustration demonstrating how the CameraLearning system sends notifications in a messaging application when detecting unfamiliar faces. Each alert includes the date and time the detection occurred, as well as a snapshot of the individual in question. This real-time information allows the homeowner to promptly evaluate any potential threat and take appropriate action.

Figure 7. Example of the Camera Learning bot interface in a messaging application, displaying photos of unknown individuals along with the detection timestamp

During testing, we noticed some variation in recognition performance across different demographic groups. In particular, the system was slightly less accurate for elderly faces and for faces with certain ethnic features that were underrepresented in the training data. Further analysis revealed that these discrepancies may be caused by an underrepresented range of facial feature appearances in the training dataset, combined with the inherent bias of the algorithm favoring the characteristics of the majority of training faces. We measured that the error rate for these specific underrepresented groups was about 12% higher than the average error rate of the system. This means that while overall accuracy was 94%, for these groups the accuracy dropped to approximately 82%, as illustrated in Figure 8.

Figure 8. Accuracy of the system on the overall test set vs. a specific subset of faces for which performance was lower. The system achieves 94% accuracy on average (blue bar), but for an underrepresented subset of the population, accuracy is about 82% (red bar), reflecting a higher error rate for that group

The results of our implementation show that the CameraLearning system is a viable solution for intelligent home surveillance, successfully identifying human faces and providing timely alerts. Achieving a 94% recognition accuracy indicates that the approach of combining deep learning face embeddings with classic preprocessing is effective for the home security context. This performance meets and in some respects surpasses current industry standards for consumer facial recognition devices. The system’s ability to operate in real time on typical home security hardware (an edge device with limited GPU capability) demonstrates its practicality. Several points merit further discussion. First, the false alarm rate, while much lower than motion-sensor-based systems, is not zero. Occasional false positives (alerts for known individuals) were observed, particularly for cases where the facial image quality was poor or the individual’s appearance changed (e.g., due to lighting or accessory changes). Fine-tuning the alert threshold and continuously updating the known faces database can help mitigate these issues. For instance, implementing an adaptive learning module that updates a known individual’s embedding over time (to account for natural appearance variations) could improve recognition consistency. Second, the bias in performance between demographic groups is a concern that reflects broader issues in AI and facial recognition technology​. Techniques such as KD-trees or locality-sensitive hashing could speed up similarity search in the embedding space if needed​.

It is also worth noting that the approach of using facial landmarks (68-point model) in conjunction with deep learning proved useful not just for performance, but for transparency. We found that being able to visualize facial keypoints and distances was helpful when explaining the system’s decisions to users. For example, if the system misclassified someone, inspecting the landmarks could sometimes reveal that poor alignment (perhaps due to an unusual head tilt) contributed to the error. This kind of insight is valuable for debugging and for increasing user trust in the system’s operation. Finally, while our system currently focuses on static facial images from video frames, an extension to utilize temporal information (multiple frames) could further improve accuracy and confidence. Tracking a face over a few frames and aggregating the recognition results can reduce uncertainty (e.g., require consistent recognition in several consecutive frames before confirming an identity). This could be implemented to filter out one-off false recognitions. Moreover, integrating motion detection to first trigger face recognition only when motion is detected in the scene can optimize resource usage, as the system would be idle when there is no activity.

Conclusions

The study presented herein has successfully implemented and evaluated a facial recognition system that achieved an accuracy rate of 94% with the XYZ dataset. This marks a notable advancement over previous systems and establishes a solid foundation for further refinement and development. This study demonstrated the effectiveness of our advanced facial recognition system, which achieved an ultimate accuracy of 85% compared to the standard XYZ dataset. Future work will focus on solving these problems by exploring the use of artificial intelligence techniques such as deep learning, which could improve the adaptability and accuracy of the system in difficult conditions. In addition, expanding the dataset to include a wider range of facial features and expressions can further enhance the reliability of the system and its applicability to real-world applications.

In conclusion, despite the fact that the existing system represents a significant advance in facial recognition technology, the pursuit of excellence in this area continues. We believe that the results of this study will make a significant contribution to the further development of more accurate and reliable facial recognition systems.

References:

1.  I. C. Society, I. of Electrical, and E. Engineers, 2019 International Conference on Sustainable Technologies for Industry 4.0 (STI) : 24-25 December, Dhaka. IEEE, 2019.
2. I. of Electrical, E. E. X. Section, I. of Electrical, and E. Engineers, 2018 IEEE In- ternational Conference on Signal Processing, Communications and Computing (IC- SPCC) : 14-16 Sept. 2018. IEEE, 2018.
3. A. International Conference on Control, 2018 15th International Conference on Con- trol, Automation, Robotics and Vision (ICARCV). IEEE, 2018.
4. S. Shaposhnikov, S.-P. gosudarstvennyi elektrotekhnicheskii universitet ”LETI”, N. issledovatelskii universitet ”MIET” (Russia), I. of Electrical, E. E. R. N. W. Section., I. of Electrical, E. E. R. Section., I. of Electrical, and E. Engineers., Pro- ceedings of the 2019 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (ElConRus) : January 28-30, 2019, St. Petersburg and Moscow, Russia, 2019. IEEE, 2019.
5. A. E. Hibaoui, M. Essaaidi, Y. Zaz, I. of Electrical, and E. Engineers, Proceedings of 2018 6th International Conference on Multimedia Computing and Systems (ICMCS). IEEE, 2018.
6. Y. Lin and H. Xie, “Face gender recognition based on face recogni- tion feature vectors.” IEEE, 9 2020, pp. 162–166. [Online]. Available: https://ieeexplore.ieee.org/document/9236905/
7. T.-H. Meen, F. Polytechnic, I. I. of Knowledge Innovation, Invention, I. of Electrical, and E. Engineers, Intelligent Applied Systems on Engineering : 2019 IEEE Inter- national Conference of Intelligent Applied Systems on Engineering (IEEE ICIASE 2019) : April 26-29, 2019, Fuzhou, Fujian, China. IEEE, 2019.
8. V. V. Das, G. Thomas, and F. L. Gaol, “Ccis 147 - face recognition system using discrete wavelet transform and fast pca,” pp. 13–18, 2011.
9. G. Guo and N. Zhang, “What is the challenge for deep learning in unconstrained face recognition?” IEEE, 5 2018, pp. 436–442. [Online]. Available: https://ieeexplore.ieee.org/document/8373863/
10. Z. Luo, J. Hu, W. Deng, and H. Shen, “Deep unsupervised domain adaptation for face recognition.” IEEE, 5 2018, pp. 453–457. [Online]. Available: https://ieeexplore.ieee.org/document/8373866/
11. Taigman Y. et al., “DeepFace: Closing the Gap to Human‑Level Face Verification,” CVPR 2014.
12. Schroff F., Kalenichenko D., Philbin J., “FaceNet: A Unified Embedding for Face Recognition and Clustering,” CVPR 2015.
13. Parkhi O. M., Vedaldi A., Zisserman A., “Deep Face Recognition,” BMVC 2015.
14. Deng J. et al., “ArcFace: Additive Angular Margin Loss for Deep Face Recognition,” CVPR 2019.
15. Wang H. et al., “MagFace: A Universal Representation for Face Recognition,” CVPR 2021.
16. Hunt M., “Burglary Statistics 2024,” Bankrate, 12 Jan 2024.​
17. Council on Criminal Justice, Crime Trends in U.S. Cities: Year‑End 2024 Update, Jan 2025.​counciloncj.org
18. Staff Report, “Crime Rate in Kazakhstan Declines by 15.2 %,” The Astana Times, 14 Sep 2023.​
19. Numbeo, “Crime in Almaty, Kazakhstan,” December 2024 snapshot.​
20. Knoema, “Kazakhstan – Burglary and House‑Breaking Rate Dataset,” 2011 entry.​
21. MacroTrends, “Kazakhstan Crime Rate & Statistics (1990‑2025),” accessed 22 Apr 2025.​
22. ZealousAdvocate, “2025 Burglary Statistics: How to Protect Your Home and Belongings,” Mar 2025.​