HUMAN BODY POSE DETECTION USING DEEP LEARNING FOR SMART HOME SURVEILLANCE SYSTEM

• Main Author: Miqdad Nadra, Muhammad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/64072
• Institution: Institut Teknologi Bandung

Developing countries cannot be separated from high crime rates, data from the Central Statistics Agency (BPS) in 2018 the percentage of the population of victims of theft crimes by the province in Indonesia is at 84.48%, this figure continues to increase every year based on BPS data in 2020 the percentage the population of victims of theft by the province in Indonesia rose to 86.51% which makes this even more worrying. The increase in crime is not followed by the readiness of the community to improve the security system to anticipate these crimes. In this final project, a residential surveillance system is created as part of a suspicious action detection system that is integrated with a cellular application using an external camera based on human body pose detection with the deep learning method of the Convolutional Neural Network (CNN). The system works by detecting the image captured from the camera, classifying the type of movement, determining the status of the detected motion, and providing notifications that are forwarded using the concept of data communication to the user when a dangerous situation is detected by the system through a mobile application. Based on the experiments and tests carried out, the dataset was created with a total data of 6000 samples which were divided into 6 movements with the division into 2 statuses, namely safe and unsafe. The model's performance after being tested resulted in an accuracy of 96.8%, error 3.2%, precision 90.4%, recall 90.4%, and f1-score 90.4% which indicated that the model was good and there was no tendency for the model to classify to a certain class. In the movement status, a simple system has also been successfully built with the addition of if and else thresholds with the status output for each movement by their respective classes. Overall the dataset generation system, movement type classification system, and movement status classification system can meet the expected specifications.