| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 5.32 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
A indústria das máquinas de venda automática é um importante contribuinte para a economia
global, sendo um setor com um crescimento notável devido à crescente procura por opções
convenientes de autoatendimento e de avanços tecnológicos, como sistemas de pagamento
sem dinheiro e máquinas de venda automática inteligentes. Apesar de serem projetadas com
características antirroubo, o roubo destas máquinas é uma ocorrência frequente, deixando os
proprietários com elevadas despesas.
Considerando a relevância da indústria de máquinas de venda automática e os desafios de
segurança enfrentados, foi tomada a decisão de desenvolver um sistema capaz de identificar
características físicas e emocionais de potenciais ladrões, com o objetivo de criar um detetor
de atividades suspeitas. Este sistema visa fornecer uma camada adicional de proteção e
mitigação de roubos nas máquinas de venda automática da Reckon.ai, mas pode ser utilizado
em qualquer propriedade, reduzindo as despesas e prejuízos enfrentados pelos seus
proprietários.
A solução proposta divide-se em três blocos fundamentais, cada um dedicado a uma tarefa
específica. O primeiro consiste numa neural network dedicada à identificação de expressões
faciais, o segundo foca-se no reconhecimento de objetos, e o terceiro aborda a deteção de
ações agressivas.
A aplicação consiste em captar imagens em tempo real a partir da webcam, verificar se existe
movimento e rostos no frame, extrair o rosto do utilizador, classificar qual a expressão facial
que este manifesta, identificar objetos que representem perigo e classificar a ação. Por fim, o
sistema analisa as respostas dos três classificadores, identificando se o indivíduo presente em
cena é ou não suspeito.
Foram empregues diversos métodos e conduzidas várias experiências com o intuito de
otimizar o desempenho dos modelos. Estas abordagens incluíram o ajuste dos
hiperparâmetros, a análise dos resultados de várias vertentes da arquitetura e a exploração
de diferentes arquiteturas. Também é feita uma comparação entre as experiências
desenvolvidas com outros trabalhos que tenham o mesmo objetivo.
O sistema final foi testado com um dataset de 200 vídeos e teve como resultado 127
avaliações corretas num total de 150 para suspeitos e 39 avaliações corretas num total de 50
para não suspeitos, representando isto uma accuracy de 0,83.
The vending machine industry is a significant contributor to the global economy, representing a sector with remarkable growth due to the increasing demand for convenient self-service options and technological advancements such as cashless payment systems and smart vending machines. Despite being designed with anti-theft features, the theft of these machines remains a frequent occurrence, leaving owners with high expenses. Considering the relevance of the vending machine industry and the security challenges faced, the decision was made to develop a system capable of identifying physical and emotional characteristics of potential thieves with the goal of creating a suspicious activity detector. This system aims to provide an additional layer of protection and mitigation of thefts in Reckon.ai vending machines but can be utilized in any property, reducing expenses and losses for its owners. The proposed solution is divided into three fundamental blocks, each dedicated to a specific task. The first block consists of a neural network dedicated to facial expression identification, the second focuses on object recognition, and the third addresses the detection of aggressive actions. The application involves capturing real-time images from the webcam, checking for movement and faces in the frame, extracting the user's face, classifying their displayed facial expression, identifying potentially dangerous objects, and classifying the action. Finally, the system analyzes the responses from the three classifiers to determine whether the individual present in the scene is suspicious or not. Various methods were employed, and numerous experiments were conducted to optimize the performance of the models. These approaches included fine-tuning hyperparameters, analyzing results from different architectural perspectives, and exploring various architectures. Additionally, a comparison was made between the experiments carried out and other works with similar objectives. The final system was tested with a dataset of 200 videos, resulting in 127 correct assessments out of 150 for suspects and 39 correct assessments out of 50 for non-suspects, achieving an accuracy of 0.83.
The vending machine industry is a significant contributor to the global economy, representing a sector with remarkable growth due to the increasing demand for convenient self-service options and technological advancements such as cashless payment systems and smart vending machines. Despite being designed with anti-theft features, the theft of these machines remains a frequent occurrence, leaving owners with high expenses. Considering the relevance of the vending machine industry and the security challenges faced, the decision was made to develop a system capable of identifying physical and emotional characteristics of potential thieves with the goal of creating a suspicious activity detector. This system aims to provide an additional layer of protection and mitigation of thefts in Reckon.ai vending machines but can be utilized in any property, reducing expenses and losses for its owners. The proposed solution is divided into three fundamental blocks, each dedicated to a specific task. The first block consists of a neural network dedicated to facial expression identification, the second focuses on object recognition, and the third addresses the detection of aggressive actions. The application involves capturing real-time images from the webcam, checking for movement and faces in the frame, extracting the user's face, classifying their displayed facial expression, identifying potentially dangerous objects, and classifying the action. Finally, the system analyzes the responses from the three classifiers to determine whether the individual present in the scene is suspicious or not. Various methods were employed, and numerous experiments were conducted to optimize the performance of the models. These approaches included fine-tuning hyperparameters, analyzing results from different architectural perspectives, and exploring various architectures. Additionally, a comparison was made between the experiments carried out and other works with similar objectives. The final system was tested with a dataset of 200 videos, resulting in 127 correct assessments out of 150 for suspects and 39 correct assessments out of 50 for non-suspects, achieving an accuracy of 0.83.
Description
Keywords
Suspect Detection System Facial Expression Classification Object Classification Action Classification Convolutional Neural Network Deep Learning