Research On Lidar-based 3D Object Detection Algorithm For Intelligent Vehicles

The assisted driving and autonomous driving applications of intelligent vehicles are the development trend of future road traffic.The object detection serves as the core basis of perception system especially for the sake of path planning,motion prediction,collision avoidance.This study aims to lightweight single-stage 3D object detection algorithms and enhance detection performance using knowledge distillation while meeting real-time requirements.Moreover,a feature-level fusion-based multi-vehicle cooperative 3D object detection algorithm is proposed to effectively improve the accuracy of the object detection algorithms in complex driving environments.The main research contents are as follows:1、This study uses structured reparameterization to lightweight the backbone network and introduces attention mechanism to improve the performance of single-stage 3D object detection algorithms.The experimental results show that the proposed optimization module significantly improves detection accuracy,especially in the recognition performance of vehicle,pedestrian,and bicycle categories while reducing 43% of the parameters.2、This study optimizes single-stage 3D object detection algorithms using knowledge distillation to improve detection accuracy while maintaining network lightweightness.Different knowledge distillation strategies are employed,including output-based,intermediate feature-based,and label-based knowledge distillation.The experimental results show that different strategies can achieve improvement effects,even surpassing the teacher model based on two-stage networks.The detection precision of the car category with medium difficulty increases by 4.17% using knowledge distillation.3、A feature-level fusion-based multi-vehicle cooperative 3D object detection algorithm is proposed to perceive the dynamic environment accurately in complex intelligent driving environments.The proposed model not only considers the shortcomings of single-vehicle object detection models but also improves detection accuracy through multi-vehicle cooperative perception.Feature alignment module and multi-scale deformable attention module are designed for the fusion model.The experimental results demonstrate the effectiveness of the proposed model on Open COOD dataset,achieving the best detection precision compared with the state-of-the-art cooperative perception models.4、Virtual scenes are built in the CARLA simulator,and simulation test data is generated to analyze the performance of the multi-vehicle cooperative method under different locations,weather scenarios,and traffic participant distributions.The experimental results show that the proposed algorithm has certain application effects in different scenarios,indicating the importance of training sample diversity for the performance of deep learning-based object detectors.In summary,the research achievements of this study have important implications for the development of object detection technology for intelligent vehicles.