Multi-view aggregation for real-time accurate object detection of a moving camera

Object detection plays an important role on various mobile robot tasks. However, directly applying existing detectors on videos from a mobile robot will cause a sharp accuracy decline, because such videos introduce some extra difficulties on accurate detection. This paper proposes a viewpoint-based memory mechanism to handle detection performance deterioration and improve detection accuracy of the videos in real time. The mechanism positively organizes previous results from multiple viewpoints of target objects as prior knowledge to enhance detection accuracy for succeeding frames, and it is designed as an extension module of an existing image detector. In experiments, we collect testing dataset from an indoor mobile robot, and compare performance of several sole image detectors and the same detectors extended by the extension module. The result shows the mechanism module achieves 20.7% object localization rate margin in average at a cost of 18.1 ms, and the mechanism can give positive impact on various existing detectors. The result indicates the proposed method achieves good accuracy margin, has acceptable time cost, and gets a degree of universal applicability.