Toward a Safer Cities: A Signal for Help Recognition System

During the isolation period linked to the COVID-19 pandemic, there was a significant increase in domestic violence, which led to it being referred to as the “Shadow Pandemic”. In response to this problem, a Canadian foundation introduced the “Signal for Help”—a discreet hand gesture that allows individuals in danger to silently ask for help by tucking the thumb into the palm and folding the fingers over it. This signal has become an essential tool for victims of abuse, as it allows them to discreetly seek help in situations where speaking out is not an option. However, its effectiveness depends on the ability of those around to recognize it and respond quickly. My thesis focuses precisely on this challenge: developing a system capable of automatically detecting the “Signal for Help” gesture in real time, using computer vision and machine learning technologies. The goal is to ensure that no cry for help goes unnoticed, thereby increasing the effectiveness of the gesture and improving the chances of timely intervention in emergency situations. This project is the development of the first functional prototype capable of detecting the “Signal for Help” gesture with an accuracy of 91% on the test set. The proposed system employs a two-stage pipeline that combines hand tracking and feature extraction, followed by real-time gesture detection. In the first stage, hands are detected and tracked using a combination of the Google MediaPipe framework and the DeepSORT algorithm, which assigns a unique ID to each detected hand. Simultaneously, hand landmark features are extracted for each frame using MediaPipe. This parallel process ensures both accurate hand tracking and real-time feature extraction, which are essential for reliably detecting the “Signal for Help” gesture. Accurate tracking is crucial because the “Signal for Help” gesture is a specific sequence of hand movements over time. Therefore, the system must ensure that each hand is properly tracked across multiple frames, without confusing one hand with another. Once a hand has been tracked for at least 20 consecutive frames, the extracted features are passed to the second stage of the pipeline. In the second stage, the features are processed by a neural network specifically trained with a dedicated dataset to recognize the gesture. If the “Signal for Help” is detected, the system automatically triggers an alarm, sending a 10-second video clip to security operators via a mobile app. The operators receive the notification and can verify the validity of the help request, taking appropriate action if necessary. This solution represents a significant advancement in the field of intelligent surveillance, improving the ability to detect help requests in real time and enabling a rapid emergency response. The system is designed not only to be highly efficient but also to minimize false positives as much as possible, ensuring a reliable and accurate detection of critical gestures. By combining advanced tracking algorithms and neural networks specifically designed for this task, the system provides a practical and efficient solution that enhances real-time gesture recognition while ensuring a high level of reliability and security in detecting critical situations.