Fast Depth Estimation of Object via Neural Network Perspective Projection

In autonomous driving and mobile robotic systems, obtaining the depths of objects in real-time is crucial. The current network-based methods usually design complex network to achieve 3D object detection or monocular depth estimation for the whole image, resulting in too slow to be applied to mobile robots. The perspective projection-based method can achieve real-time, which calculates the object depth based on the camera parameters, the object sizes in the world coordinates and in image coordinates. While it relies heavily on the accuracy of object size in images coordinates, and the size is usually obtained with errors through detector network. Combining the perspective projection-based methods and network-based methods, we propose a fast object depth estimation method by designing a neural network to learn perspective projection, called Fast-Depth-NPP: 1) Instead of considering the whole image, we only consider the local depth of the image; 2) Using local image patches as network inputs avoids measurement errors of object size with detector; 3) the use of global information is enhanced by incorporating position encoding. Our method is validated on the mobile robot public dataset Neurons Perception dataset, achieving excellent results and meeting the real-time requirements.