Object Distance Estimation in Monocular Image using Deep Neural Network

Abstract

Autonomous driving is become a hot topic in computer vision, Robotics tasks. It is divided into cognition, judgment, and control. In Cognition part, robot, vehicle are sensing circumstance using mounted their vision, recognition sensor. So, Judgment, Control are determined use its information. The information is exist many direction like velocity of agent, location of object and so many dynamic environment. Commonly, variety sensor use for Autonomous driving as GPU, Camera sensor, IMU, LiDAR. However most of sensors are high cost and demand high computation. Like mostly we need relatively low cost and low computation, so infer smaller device. The stereo camera it is. Stereo vision is copy to human’s eyes, and they can making disparity map at difference between one eye and the other eye. It is inversely proportional depth map as relative distance of objects in image. This paper based on this theory. But, stereo camera using uncommonly in real environment, so we have monocular camera like black box. Black box video is commonly approximated by media and many circumstance. However using only one eye can’t make disparity from the other one, we need geometry of image. Geometry is connected tightly camera view point called Camera Pose. The Camera pose is moving view point by agent, mostly using location estimation and 3D projection from 2D point. Conducting two task simultaneously called SLAM (Simultaneous Localization And Mapping) in more detail Visual SLAM. Knowing location of agent in map and draw surrounded objects are very difficult but important task for autonomous driving. Moreover, only using monocular camera make problem more difficult. Also unlike overseas like America, or Germany, South Korea has high density of objects on the road as building, fence, pedestrian, this make hard to calculate geometry of image. So, traditional algorithms have limitations, we propose using Deep Neural Network. Deep learning widely use every task, optimized to limitation of traditional calculation at image processing algorithm. They accept inherent architecture in image, intensity, brightness contrast, everything of image. So, The contribution of this paper, first, we only use monocular camera as black box video has high density in South Korea, Second, we train to Deep Neural Network but they don’t have any ground-truth, Third, for training more complex architecture in image, using pre-trained Semantic Segmentation task