Single shot detector for object detection using an ensemble of deep learning and statistical modelling for robot learning applications

Abstract

In this study, a deep learning model was synthesized for object detection that can give better results than the state-of-the-art accuracy and reliability in production environments making it more accessible to production-ready applications. For this, a model architecture was used that is inspired by state-of-the-art Single Shot Detector architecture. Statistical analysis was performed on the results generated by the model to make final predictions in each frame. Experimental results outperform simple-SSD architecture. For 300 X 300 inputs, it achieves 84.7% mean Average Precision on the system with Nvidia GeForce GTX 1660 Ti 14GB, 16 GB ram, and Intel(R) Core (TM) i7-9750H CPU @ 2.60GHz at 60FPS, outperforming existing architectures of SSD, Faster Region - Convolutional Neural Network, You Only Look Once (YOLO), Resnet50, Resnet100 in use.