Research On 3D Object Detection Algorithm For Autonomous Driving In Adverse Weather

With the growing maturity of autonomous driving technology,intelligent vehicles equipped with different levels of autonomous driving technology are gradually coming into the public eyes.Compared to other intelligent terminals,the driving safety of intelligent vehicles is particularly important because it is closely related to our personal safety.The environment perception system is the eyes and ears of the intelligent vehicle,which can sense the environment around the vehicle in all directions and assist in subsequent decision making and tracking control systems;3D object detection is one of the most important tasks in the environment perception system,which captures information about other traffic participants in the 3D space in real time,and assisting in other perception tasks.However,under various adverse weather conditions such as rain,snow and fog,the performance of the 3D object detection models decreases significantly,which has a serious impact on other tasks and poses a threat to the driving safety of intelligent vehicles.This paper investigates how to improve the performance of 3D object detection models in autonomous driving scenarios under adverse weather conditions from both the model and data perspectives.From the model perspective,this paper makes use of the redundancy of environmental information brought by multimodal data fusion to compensate for the distorted information of single-modal point cloud data,and constructs a 3D object detection post-processing optimisation model based on the existing excellent point cloud detection model and the highly compatible fusion detection framework;from the data perspective,this paper covers two mainstream data processing strategies for adverse weather conditions,and combines the 3D object detection models with data augmentation methods in the training stage and data preprocessing methods in the testing stage in order to achieve further model performance gains.The main research contents of this paper include:(1)Structural optimization design of the 3D object detection modelsThis paper references representative models of existing point cloud-based and multimodal fusion-based 3D object detection,and constructs the 3D object detection post-processing optimisation model by combining the existing excellent point cloudbased 3D detection model and the highly compatible post-fusion detection framework.Further,the post-processing optimisation model and the baseline model are trained with the existing open source dataset,and the performance gain of the post-processing optimisation model over the baseline model is tested by the qualitative and quantitative experiments.(2)Data augmentation methods for adverse weather in the model training stageThis paper introduces several data augmentation methods for adverse weather in the model training stage,which cover different types of adverse weather effects on multi-modal sensor data,and constructs a comprehensive adverse weather dataset with the cross-weather and cross-modal asymmetric distortion data augmentation strategy proposed in this paper.This paper trains the standard 3D object detection models with the dataset constructed itself in order to improve the robustness of the models to the real adverse weather.(3)Data pre-processing methods for adverse weather in the model testing stageThis paper introduces several pre-processing methods for adverse weather data in the model testing stage,targeting both multiple sensor data modalities and different adverse weather types.This paper improves the semantic segmentation-based point cloud denoising algorithm and employ the learning-based image restoration algorithms.This paper pre-processes the distorted sensor data before the sensor data stream enters the 3D object detection models in order to improve the performance of the 3D object detection models under real adverse weather conditions.(4)Implementation of the 3D object detection models for adverse weatherBased on the existing high-performance point cloud 3D object detection baseline model,this paper constructs several 3D object detection models for adverse weather throuth ablation experiments by combining the post-processing optimisation method for the model network structure,the data augmentation methods in the model training stage and the data pre-processing methods in the model testing stage.Further,all the models constructed in this paper were trained based on the existing hardware and software conditions,and the model performance was tested and compared under real adverse weather conditions by the quantitative and qualitative experiments,which can verify the effectiveness of the different optimisation methods and the actual 3D object detection performance of the optimised models.