SPORTS ITEM DETECTION USING MOBILENETV2 WITH SINGLE SHOT DETECTOR

Abstract

Object detection can be applied in various situations and systems. In this research, object detection was specifically applied to build a lending system for tracking the details of sports items in a students' dormitory. The current sports item lending system with a manually recording of data was time consuming, and are more prone to human errors. This inspires the researcher to build a real-time object detection web-based sports item lending system. The system was trained using Single Shot Detector (SSD) with Mobilenet V2 technique to detect the sports item in the warehouse. A total of 960 self-collected sports item image data was applied in four different experiments with the same batch-size configuration and learning rate value of 0.02. From the experiments, several models with different number of training iterations and training data were built to find the best model to be implemented in the sports item lending system. The best model was obtained from the second experiment with a high accuracy of 0.93 mean average precision (mAP), a confidence of 97%, and a total loss of 0.28. For future work, it is recommended to increase the volume of training data, include other variations of objects in order to further improve the results, and apply other object detection techniques for comparison purposes.

Keywords

• deep learning
• machine learning
• mobilenetv2
• object detection
• single shot detector