Object Detection For Images And Point Clouds In Autonomous Driving

The object detection algorithm acquires the category and position information of the interested objects in the driving scenarios,which is the central and fundamental part of guaranteeing the reliability of autonomous driving.However,it is difficult to achieve accuracy and reliability of prediction with single-modal object detection algorithms.Providing perception information of road scenarios jointly with different modal data is an inevitable trend in the development of autonomous driving.This paper investigates the two critical issues of reducing the cost of manual annotation and improving the integrity of object information.Novel object detection methods are explored with image and point cloud data,respectively,which improve the effects of variation in data distribution on algorithm robustness and generalization ability.The main innovative works are as follows:(1)For the purpose of improving the distribution differences resulting in pixel and semantic features,a novel semi-supervised domain adaptive object detection method based on hierarchical feature consistency constraint is proposed to explore the detection capabilities in complex road scenarios.This method constructs the consistency of pixel and semantic features in different domains with an adaptive instance normalization layer and distribution metric function.The experimental results on public benchmark datasets demonstrate that constrained pixel and semantic feature distributions assist in learning domain-consistent features,which can improve the detection ability for objects in complex road scenarios.(2)For the purpose of extending the detection capability for low-resolution objects,a semi-supervised domain adaptive object detection method with a joint attention mechanism and hierarchical feature consistency constraint is proposed.Based on the hierarchical feature consistency constraint method,this method improves the ability to refine the fine-grained features by introducing feature dependency with the self-attention mechanism.The consistency of pixel and semantic features is constructed with an adaptive instance normalization layer and cross-attention mechanism,which can enhance the cross-domain semantic relevance of foreground objects.The experimental results on the public benchmark dataset demonstrate the joint construction of feature dependencies and category correlations assist in learning the critical features,which can improve the detection ability for low-resolution objects.(3)For the purpose of reducing false-negative samples generated by the prediction biases of pseudo-labels,a novel unsupervised domain-adaptive object detection method based on conditional attention and domain-consistent directional gradient updating is proposed to explore the detection capability in unknown road scenarios.This method learns domain consistency features in the gradient updating of the detection network,which constructs the consistency of pixel features and semantic features by the domain consistency direction gradient updating method.Learning the critical features of the object in the discriminative network increases the distinguishing capability of domain categories with the conditional channel attention module.The experimental results on the public benchmark dataset demonstrate this method contributes to increasing the detection ability on unknown road scenarios,which provides rich semantic information for objects.(4)For the purpose of exploring the prediction capability for spatial location information,while predicting the color and texture information,a novel object detection method based on point cloud and neighbor alignment mechanism of graph neural network is proposed.This method introduces a neighbor alignment mechanism in graph neural networks,which can enhance the learning ability of point cloud representation by reducing the sensitivity of the input variance.The experimental results on the public benchmark dataset demonstrate the graph neural network based on the neighbor alignment mechanism assists in improving the integrity of object information,which provides accurate spatial location information for road scenarios.In summary,the detection results of image and point cloud data are jointly used to provide the color,texture,and spatial location information of the road scene,which contributes to increasing the tolerance and reliability of the object detection algorithms.This method provides the fundamental research for advancing the commercial application in high-level autonomous driving.