DESIGN AND IMPLEMENTATION PEOPLE COUNTING SYSTEM USING DETECTION AND TRACK DEEP LEARNING BASED

• Main Author: Cahyadi, Nanang
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/64380
• Institution: Institut Teknologi Bandung

The importance of statistic object counting in human behavior analysis drives efficiency improvement in the data collection process. Main problem that happened is manual counting by human efforts. Manual counting can be caused into low accuracy as a result of human limitations. This research study people counting system built by detection and track approach by deep learning based that aims to automate counting process by analyze recorded CCTV video that prioritize good accuracy. The research is done by build human detector with its dataset, then integrate track module and counting method. Building human detector includes images gathering and object annotating for 4000 images that contains 2000 that from extracted CCTV video and taken from Open Image Dataset V6 for the rest. Those images feed into YOLOV4 and reach mAP 92.62%. The integration process is carried out the module using DeepSORT for the trajectory of the detected object movement is obtained. Then the counting method used is to register the movement of objects seen from the difference in the location of the detected trajectory with the location of the last detected object. The method will give the results of counting objects that come out and enter in the video. In order to get optimal results, tuning is done in the DeepSORT parameter and the counting method. The system is tested using the CCTV video that is owned where the scenario being tested is that the movement of humans in the video is two-way, the F-score value for counting objects that come out and enter is 100% and 85%, respectively.