Project

Design and Implementation of a Deep Neural Network based Real Time Object Detection for GoPiGo

In this paper we present a new neural network architecture, MobileNet-Tiny that can be used to harness the power of GPU based real-time object detection in raspberry-pi and also in devices with the absence of a GPU / restricted graphic processing capabilities such as mobile phones, laptops, etc. MobileNet-Tiny trained on COCO dataset running on a non-GPU dell xps 13 laptop, achieves an accuracy of 19.0 mAP and a speed of 19.4 FPS which is 3 times as fast as MobileNetV2, and when running on a raspberrypi it achieves a speed of 4.5 FPS which is up to 7 times faster than MobileNetV2. Also, model size of MobileNet-Tiny is 4.5MB which is 5 times smaller than MobileNetV2. MobileNet-Tiny was modeled to offer a compact, quick and well balanced object detection solution to variety of GPU restricted devices.

Items in ScholarWorks are protected by copyright, with all rights reserved, unless otherwise indicated.