Introduction:
During my internship at Singapore University of Technology and Design, I had the opportunity to work on a fascinating project centered around action recognition under the Assoc Prof Yuen Chau. This project aimed to address a significant issue in housing schemes in Singapore, particularly in HDBs (Housing Development Board) where the ground floor, known as the void deck, serves as a communal area. The challenge was to detect and prevent unauthorized activities such as football playing and cycling, which are prohibited to ensure cleanliness and safety. Leveraging machine vision and deep learning techniques, our team developed an accurate and lightweight solution that could run on edge devices. In this article, I will share the insights and accomplishments gained throughout this project.
Problem Background:
HDBs and Condos are popular housing schemes in Singapore, with HDBs being more affordable options provided by the government. The void deck, located on the ground floor of HDBs, serves as a common area for residents to socialize and relax. However, unauthorized activities like football playing and cycling posed challenges to maintaining cleanliness and safety within these spaces.
Objectives:
The primary objective of our internship project was to develop a system that could automatically detect and recognize unauthorized actions within the void deck using machine vision and deep learning techniques. We aimed to create an accurate and lightweight solution that could be deployed on edge devices, ensuring real-time monitoring and immediate intervention when necessary.
Methodology:
During this first internship project, we followed a rigorous methodology to develop an effective solution for detecting unauthorized actions within the void deck of HDBs. Given the responsible nature of the project, specific details regarding the codes and algorithms used cannot be disclosed. However, I can provide an overview of the general approach we undertook.
1. Problem Analysis: We began by thoroughly understanding the problem statement and the objectives of the project. This involved studying the rules and regulations of HDBs regarding unauthorized activities within the void deck.
2. Dataset Collection: To train our action recognition system, we collected a comprehensive dataset of annotated videos showcasing various actions within the void deck. This dataset was crucial for training and evaluating the deep learning models.
3. Model Selection and Experimentation: We explored various types of deep learning models suitable for action recognition tasks. These included convolutional neural networks (CNNs), recurrent neural networks (RNNs), and their combinations such as CNN-LSTM architectures. By experimenting with different models, we aimed to identify the most effective approach for our specific problem.
4. Training and Evaluation: Using the collected dataset, we trained the chosen deep learning models. This involved feeding the videos into the models and fine-tuning the parameters to optimize performance. We evaluated the models based on metrics such as accuracy, precision, recall, and F1-score.
5. Optimization for Edge Devices: Given the requirement for a lightweight solution capable of running on edge devices, we focused on optimizing the selected model. This optimization process involved techniques such as model compression, quantization, and pruning to reduce the model's size and computational complexity.
6. System Integration and Testing: Once the optimized model was ready, we integrated it into a complete system capable of real-time action recognition. We thoroughly tested the system, ensuring its accuracy, speed, and reliability.
Results and Achievements:
Through extensive experimentation and evaluation, we successfully developed a robust action recognition system for detecting unauthorized actions within the void deck of HDBs. Although specific details regarding the chosen model and algorithm cannot be disclosed, our approach involved trying various deep learning models before settling on a solution that provided accurate results.
Conclusion:
The internship project on action recognition allowed us to gain valuable insights and experience in developing real-world solutions using deep learning techniques. By experimenting with different models and optimizing for lightweight deployment on edge devices, we achieved a reliable system for detecting unauthorized actions within the void deck of HDBs. This project not only enhanced our technical skills but also emphasized the importance of problem-solving and adapting algorithms to specific requirements.
Comments