Please use this identifier to cite or link to this item: https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4537
Title: Optimized Real-time object detection on Mobile phones
Authors: Rajeevan, Kuganathan
Issue Date: 12-Aug-2021
Abstract: The usefulness and feasibility of optimized object detection models in real time applications for mobile phone and embedded device platforms is analyzed in the research. Such models will be useful as the number of mobile devices and IoT devices and the applications are widely increasing. The problem in mobile hardware platforms is that the CPU is not designed to do heavy operations as PC hardware to conserve power. To achieve real-time detection it is required for a different model or optimizing and improving the existing model. The objective of the research is to find a solution which gives real-time performance without compromising accuracy of the results on mobile phones. Existing object detection techniques were analyzed including Deterministic approaches and Machine learning models. Through literature review it is found that YOLO machine learning model can be improved to make it suitable for mobile hardware. Various techniques applied and tested including model trimming, add redundant layers, changing layer sizes and use different tools to compress the trained model. In the results it is found that combining such techniques improve the performance while not sacrificing much accuracy. Tests were done on Mobile Phone and PC hardware with same data set and compared. And also test with different models on Mobile phone also compared. The results show around 8 times improvement on inference time on mobile devices than using base models. A proof of concept application created by training the street sign data and used in android camera application to detect street signs on real time. The results show significant effectiveness on optimized object detection models on mobile phones.
URI: http://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4537
Appears in Collections:2020

Files in This Item:
File Description SizeFormat 
2017 MCS 062.pdf456.46 kBAdobe PDFView/Open


Items in UCSC Digital Library are protected by copyright, with all rights reserved, unless otherwise indicated.