VMM: Viewpoint-based Memory Mechanism for Object Detection of Moving Sensors

Abstract

Video object detection is a promising technology to enhance perception capability of mobile intelligent devices in practical scenes. Recent algorithms continously achieve remarkable detection performance on popular datasets. However, the performance is still unsatisfactory on practical scene videos from a moving sensor. In this paper, we propose a viewpoint-based memory framework to improve detection performance by exploiting viewpoint information of the sensors, which obviously enhances the detection accuracy at a minor cost of execution time. Videos and the sensors’ corresponding angle information are collected as test dataset by a camera-IMU module, and the memory framework is actualized as an extension module of a mainstream image detector. Experiments are designed to compare performance of the image detector and the framework extended detector. The experiment results indicate the extended detector achieves a 25.0% average localization margin and costs extra 12.36 ms/frame in average compared with the image detector.