3D Gaussian-Driven Depth Estimation for Efficient and Accurate Agricultural Produce Picking

Fruit and vegetable picking robots need accurate depth information to localize fruits, but traditional methods such as depth cameras are costly, binocular cameras have limited accuracy in complex scenes and are cumbersome to operate.In this paper, a self-supervised depth estimation method based on monocular camera and 3D Gaussian is proposed for fruit and vegetable picking robots.The method utilizes 3D Gaussian to generate left, center, and right viewpoint images from the images captured by the monocular camera, and transforms them into disparity maps through their intrinsic depth information, which are used as supervisory signals for the disparity network.The RAFT-based network architecture is utilized for disparity estimation and optimized by combining photometric loss and disparity loss.The experimental results show that the method can effectively reduce the cost of data acquisition, improve the accuracy and robustness of disparity estimation, and has a broad application prospect in practical applications such as fruit and vegetable picking.