OBJECT DETECTION METHOD IN EMBEDDED SYSTEM FOR REAL-TIME OBJECT TRACKING ROBOT CONTROL PROBLEM

Abstract

In this paper, we conduct a comparative study of the original SSD neural network architecture using VGG-16 as the backbone network with a modified SSD
architecture by replacing the VGG-16 backbone network with different versions of the MobileNet network. The goal of the study is to build an optimal deep
convolutional neural network model that ensures a balance between accuracy and speed in the object detection and tracking problem so that it can be executed
on an embedded device platform with limited computational resources. The models are evaluated on a Jetson Nano for datasets of different sizes and
complexities to have a comprehensive conclusion about accuracy and speed. The proposed method based on MobileNet network achieved almost equivalent
accuracy and achieved much faster inference speed than the original SDD model using VGG-16 network, specifically achieving the highest overall mAP accuracy
of 84% on the test dataset and an average inference speed of ~25 FPS after optimization.