DTB-Net: A Detection and Tracking Balanced Network for Fast Video Object Detection in Embedded Mobile Devices

Abstract

Recently, object detection is of great significance and full of challenges in the task of video analysis with the continuous breakthrough of convolutional neural networks. However, detecting video objects in embedded mobile devices still remains challenging. The two major reasons limit the video detection performance of existing methods in embedded mobile devices: 1) The computing capability of embedded mobile devices is limited; 2) The motion and change of the interested object causes a lot of noise and blur. To address these problems, we propose a novel method for fast video object detection in embedded mobile devices, called a Detection and Tracking Balanced Network (DTB-Net). DTB-Net performs an accurate and fast object detection in embedded mobile devices by a balance module to regulate the working state of tracker and decide on the final output of network. In addition, we build a new Robot Detection Dataset for video object detection, called the FIST -RD. A series of detailed evaluation experiments in Nvidia Jetson TX2 platform based on FIST-RD dataset demonstrate the competitive performance of proposed DTB-Net.