Research On Road Perception Based On Multi-modal Information Fusion For Vertical Control

With the development of intelligent connected vehicles and the continuous iteration of controllable suspension systems,utilizing the rich sensors on intelligent connected vehicles to improve vehicle ride comfort has attracted widespread attention from the industry and academia.For vertical control oriented road surface perception,it is necessary to perceive and reconstruct the road surface features in front of the vehicle,with a complex working environment and high accuracy requirements.Road surface perception systems based on a single sensor often only obtain a single modal information.It is difficult to meet the robustness and generalization requirements of various application environments.Therefore,multi-source heterogeneous information fusion perception technology is of great significance to exploring the potential of intelligent suspension control,improving chassis performance,and enhancing vehicle riding comfort.This dissertation is proposed and developed in this context,focusing on solving three problems.The first is to study the multimodal 3D detection method for road surface targets in response to the limitation of semantic information in road surface perception systems.The second is to study the vehicle pose estimation method and road elevation map reconstruction method based on the motion uncertainty of pose estimation.The third is to validate the application of the suspension preview control oriented road perception system for effectively utilizing the obtained preview information.And the main research content of this dissertation is as follows:Firstly,a multi-source heterogeneous information fusion road perception system was built.And the detailed hardware configuration of the system platform was provided.The characteristics of hardware such as monocular cameras,MEMS solid-state Li DAR,and IMU were analyzed.Then,the principles of camera Li DAR and Li DAR-IMU joint calibration were modeled.At the same time,the frequency characteristics of data transmission between various sensors were analyzed.Consequently,a method of hardware time synchronization and software time synchronization was adopted to synchronize data between sensors.Moreover,the mechanism of image distortion and Li DAR motion distortion was analyzed,and the preprocessing methods required for each sensor were studied.Finally,the software architecture of the multi-source heterogeneous information fusion road perception system was provided.Secondly,to address the problem of information loss in point cloud feature extraction,an importance sparse convolution module was designed,which incorporated the filtering of the importance of the input features.It effectively alleviated the loss of geometric feature description ability due to the feature collapse of the sparse convolution features.Based on this,a Li DAR single mode 3D object detection method was constructed.On this basis,a 3D road target detection method based on the multimodal complementary attention mechanism was proposed.The proposed method efficiently captured the correlation between images and point clouds to achieve effective alignment of heterogeneous information.And it improved the fusion degree of the two types of features between point cloud and image.In addition,considering that comfort-sensitive targets such as speed bumps and manhole covers were not easy to be detected with small sizes,a multi-scale image feature generation layer was established.The receptive field block was used to efficiently construct multi-scale features.And the adaptive spatial feature fusion was used to generate feature maps for deep and shallow feature aggregation.Finally,validation was completed on the dataset established through data collected from the real vehicle.The results of comparative experiments and ablation experiments showed that the algorithm proposed in this dissertation had good detection accuracy and met real-time requirements.Thirdly,based on the concept of dimensionality reduction,a down sampling method was used to reduce the number of original point clouds obtained from MEMS solid-state Li DAR while ensuring the geometric features of the original point clouds.The local smoothness was used to select the required feature points for vehicle pose solving.And combining the characteristics of MEMS Li DAR to form the smoothness of reflection intensity to alleviate the problem of small field of view angles.Further,real time solution of vehicle pose was achieved by establishing constraint equations for feature points and minimizing feature residuals.After obtaining accurate pose,an optimal mapping algorithm for local elevation in front of vehicle based on motion uncertainty was proposed.It introduced a Li DAR sensor noise model to construct a road grid height model.And a unified description of grid height measurement was formed through maximum likelihood estimation.Then,the map was updated by one-dimensional Kalman filter.Moreover,it designed a map update strategy considering motion uncertainty,which derived an error propagation method for introducing uncertainty into vehicle pose estimation.Meanwhile,weighting the motion uncertainty in the horizontal direction through confidence ellipse theory.In addition,due to the lack of physical laws to describe the changes in elevation,a gate strategy based on Mahalanobis distance was adopted to filter the measurements that fall into the grid to cope with scenarios of rapid elevation changes.Finally,real vehicle experiments were conducted to validate the pose estimation algorithm.And the results showed that the proposed pose estimation algorithm can accurately estimate the vehicle pose and correct errors in a timely manner when the system drifts.Furthermore,the accuracy and real-time performance of the elevation estimation algorithm proposed in this dissertation were verified through vehicle experiments using test cuboid and speed bumps with known dimensions.Finally,based on the high-precision road preview information obtained from the multisource heterogeneous information fusion road perception system,an adaptive suspension preview control algorithm based on explicit model predictive control was proposed for improving ride comfort.Focused on vertical dynamics,a quarter vehicle model considering preview information was established based on the characteristics of explicit model predictive control,which introduced elevation information as an augmented state into the suspension model.Then,in order to solve the real-time problem of the algorithm,explicit model predictive control is used to convert the original online optimization of quadratic programming into offline optimization of multi-parameter quadratic programming.Therefore,online optimization only required table lookup according to system feedback and target values,greatly reducing the amount of online computation.Moreover,in order to fully utilize various preview information,a control weight adaptive switching strategy based on preview information was constructed,which adaptive switched the optimal explicit model predictive control weight parameters under current operating conditions based on preview information.Meanwhile,it proposed an improved particle swarm optimization algorithm based on augmented Lagrange method for weight optimization under various operating conditions.During the optimization process,the suspension physical constraints in explicit model predictive control were considered,and particles were guided to move towards the optimal feasible region by continuously adjusting the multiplier and penalty factor.Finally,the algorithm proposed in this paper was compared and validated through virtual environment experiments,which demonstrating its effectiveness.