Improvement Research Of Intelligent Driving Object Detection Algorithm Based On Point Cloud And Images Fusion

With the rise of car ownership,urban traffic issues are increasingly severe,with road traffic accidents becoming more frequent.Intelligent vehicles are gradually being seen as a potential solution to urban traffic problems by assisting drivers in driving.The environmental perception module is one of the most crucial components of intelligent driving systems,as the performance of this module directly impacts the decision planning and motion control of intelligent vehicles.Furthermore,good perception performance is critical for ensuring the safe driving of intelligent vehicles.Deep learning has emerged as a new development direction for intelligent vehicle environmental perception methods in recent years.While point cloud and image fusion perception techniques have achieved some research results,they only apply to simple scenarios with suitable environmental conditions.Daily driving scenarios often involve mutual obstruction of objects,poor weather conditions,and insufficient lighting conditions,which can hinder the sensors from collecting effective information and affect the detection performance of algorithms.Therefore,achieving accurate and robust object detection in complex and variable driving scenarios is an urgent problem for intelligent vehicle fusion perception algorithms.This thesis proposes a data fusion method for intelligent vehicles and develops a 3D object detection algorithm for driving scenarios to address the problem of poor detection capability of existing typical 3D object detection algorithms in real driving,as follows:(1)Research on intelligent vehicle environmental perception systems.Given multiple input data in the current intelligent vehicle environment sensing system,we aim to efficiently develop a 3D object detection algorithm.Firstly,we analyze the performance of commonly used onboard sensors and their combinations in the environment sensing system.In this thesis,we determine the use of camera and Li DAR as the data source for the object detection algorithm.Secondly,based on the sensor arrangement scheme,we analyze the environment sensing system architecture and conclude that a sound environmental sensing system should leverage the advantages of different sensors through multi-sensor data fusion to provide comprehensive and redundant environmental information for the object detection algorithm.(2)Research on the improvement of existing data fusion methods.Firstly,we analyze the principle and fusion levels of multi-sensor data fusion and determine feature-level fusion of point cloud and images information using neural networks.Secondly,we study the critical technology of data fusion and analyze the representative 3D object detection algorithm of multi-sensor fusion to summarize the current data fusion methods.It is found that current representative fusion methods have the characteristics of relying on images or point cloud data for detection,which is challenging to adapt to the complex and changing driving scenes that result in decreased sensor information collection ability.Finally,based on the characteristics of an excellent environmental sensing system,this thesis proposes a data fusion method suitable for intelligent vehicles.(3)A 3D object detection algorithm in a traveling scene is proposed.Based on the improved data fusion method,this thesis proposes a 3D object detection algorithm in real driving scenes called PINet.The algorithm incorporates point cloud and images fusion and utilizes a jumping feature pyramid module to provide fused feature maps with deep semantic and shallow detail information for object detection at different scales.A dynamic feature map fusion module is designed to assign weights to feature elements in different data,enabling the effective fusion of information from two branches.(4)The proposed algorithm is evaluated through comparison with other representative3 D object detection algorithms,achieving a mean Average Precision of 56.8% and nu Scenes Detection Scores of 67.2%,outperforming other algorithms.Ablation experiments also verify the effectiveness of the jump feature pyramid and dynamic feature fusion modules.Furthermore,the algorithm’s detection effectiveness is qualitatively evaluated in different traffic,weather,and lighting conditions,demonstrating its advantages in real driving scenes.In summary,this thesis systematically studies the object detection algorithm for intelligent vehicles.The system characteristics and algorithm data source are identified by analyzing the environment perception system.The fusion methods for the object detection algorithm are summarized,and an improved scheme is proposed based on the characteristics of the environment perception system.The object detection algorithm is then designed accordingly,which achieves accurate and robust 3D object detection in driving scenes,providing a favorable research foundation and technical support for intelligent vehicle object detection technology.