Detecting More Objects Indoors: a Depth-guided Detector for Mobile Robots

In robotic target object manipulation tasks, object detection plays a crucial role on ensuring types and general positions of target objects. However, directly applying existing detectors in real-life scenarios is hard to achieve required performance. This is mainly because long distance and complicated environment cut down visual information quality of the target objects. To increase valid detection distance and accuracy for the robots, this paper proposes a depth-guided detector. In this detector, depth information is utilized to select a subregion of the current frame to obtain better visual information, and a motion tracker is exploited to relate adjacent frames. Our experiment shows the proposed method enhances the valid detection distance in real-life scenarios in an acceptable efficiency cost.