Deep Learning Based Object Detection Combined with Internet of Things for Remote Surveillance

Abstract

Object detection is the key process in any video surveillance application. In case of remote surveillance, it is a necessity to accurately detect the target and transmit the detected data rapidly to main station so that further actions can be taken. This paper concentrates on a framework which uses deep neural network and Internet of Things for target detection and transferring detected information to the cloud at low transmission rates. The detection framework is based on combination of YOLO-Lite which is a simpler version of you only look once (YOLO) detector and spatial pyramid pooling (SPP). When trained on COCO dataset, YOLO-Lite + SPP model runs at a speed of 40 fps with mAP of 35.7% on non-GPU platform. Performance of the same has been analyzed on PASCAL VOC, COCO, TB-50 and TB-100 dataset. On GPU based platform, precision and recall values of 89.79% and 91.67% has been achieved with processing speed of 218 fps. ThingSpeak platform has been used for data reception on cloud. Results in real-time are also demonstrated which proves the efficiency of the anticipated framework and also confirms its suitability for remote video surveillance.