Research And Implementation Of Indoor Smart Car Perception Enhancement Based On Depth Estimation

In recent years,with the rapid development of deep learning and computer vision technology,indoor smart cars have been widely used in many areas of social production and life,and gradually formed an indoor perception control with lidar as the mainstay and the cooperation of multiple sensing devices.Program.Single-line lidar is widely used in indoor robot sensing systems due to its simple structure,low cost,high accuracy,and good stability.Although single-line lidar sensing can effectively carry out map construction and positioning,single-line lidar can only obtain a single-plane sparse laser point cloud during navigation.It has insufficient perception of non-scanning plane obstacles and inaccurate perception of nonuniform rigid bodies,which is difficult to meet the complex requirements.In the scenario,there are practical requirements such as obstacle sensing and local navigation of the indoor smart car.This paper takes the space perception enhancement of the indoor smart car as the basic research object,based on the use scene of the indoor smart car,introduces the real global pose and camera internal parameters as auxiliary data for obtaining depth information;proposes a dense depth estimation method for monocular continuous RGB images based on deep learning.Use refined module design,quantitative pruning and other methods to compress the depth estimation network;speed up network inference speed and reduce performance loss.The Random Consistent Sampling(RANSAC)algorithm,combined with 2D single-line lidar data,is used for scale correction and perception fusion of dense depth data to enhance the spatial perception ability of the indoor smart car,and complete the indoor smart car perception enhancement navigation system on the embedded mobile platform.Design and implementation.The main work of this paper is as follows:(1)Indoor RGB-D data set production.This paper produces an actual indoor RGB-D data set,including more than 10,000 original RGB images in 5 actual indoor scenes,more than 5,000 sets of aligned RGB-D and pose data,40 min 60GB indoor car positioning,perception,path planning,Navigation data.Use the data enhancement method to complete the data set amplification.Provide basic data support for the following depth estimation network training and enhancement of the perception system verification.(2)Lightweight depth estimation network design.Aiming at the DeepV2D network,this paper deeply analyzes the actual scene and usage requirements,deletes the redundant pose network,analyzes the encoding-decoding network,and streamlines the depth estimation network structure.At the same time,ShuffletNet and atrous spatial pyramid pooling(ASPP)are used for lightweight network feature module design,differentiated adjustments to the encoding-decoding network structure,network pruning,and monocular depth estimation of the embedded platform The lightweight design and implementation of the network provide basic dense and deep information for the enhancement of the perception ability of the indoor smart car.(3)Perception enhancement system of deep data fusion.Analyze the existing systems and functions,carry out detailed functional module design,integrate the depth estimation network into the existing perception control system,and reduce the coupling between the system functional modules and the performance loss of the embedded mobile platform.Based on the geometric characteristics of the rigid body of the real space of the car,the depth data space perception boundary is clarified;the single-line lidar is used to jointly correct the depth estimation data and the 2D plane point cloud depth data perception fusion to enhance the space perception ability of the indoor smart car.Completed the design and implementation of the perception enhancement system based on the deep data fusion of the embedded platform.Experiments prove that the depth estimation network proposed in this paper performs well on both public and self-made data sets,and the inference acceleration is obvious on lowperformance devices.At the same time,the depth data fusion perception enhancement scheme proposed in this paper is based on depth estimation.Each functional module runs stably and efficiently on the embedded platform of the indoor smart car,which can effectively enhance the space perception ability of the indoor smart car.