{"id":{"repo_id":"msu","oai_identifier":"oai:d.lib.msu.edu:etd_52748"},"canonical_url":"https://search.ndltd.org/etd/msu/oai:d.lib.msu.edu:etd_52748","repository":{"repo_id":"msu","name":"Michigan State University","base_url":"https://d.lib.msu.edu/oai"},"display":{"title":"DOMAIN ADAPTIVE 3D OBJECT DETECTION FOR AUTONOMOUS VEHICLES","abstract":"Thesis (Ph.D.)--Michigan State University. Electrical and Computer Engineering - Doctor of Philosophy, 2025","abstract_html":"Thesis (Ph.D.)--Michigan State University. Electrical and Computer Engineering - Doctor of Philosophy, 2025","abstract_has_math":false,"creators":["Lu, Xiaohu"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":["Radha, Hayder","Morris, Daniel","Traganitis, Panagiotis","Liu, Xiaoming"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2025,"date_issued":"2025","date_published":"2025","updated_at":"2026-11-18T18:59:26Z","subjects":["Engineering","Electrical engineering","Computer engineering"],"languages":["eng"],"rights":["In Copyright"],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["etd:52748","local:Lu_grad.msu_0128D_21726"],"render_values":[{"text":"etd:52748","href":null,"code":true},{"text":"local:Lu_grad.msu_0128D_21726","href":null,"code":true}]}],"person_identifier_entries":[]},"links":{"outbound_url":"https://doi.org/doi:10.25335/zzrr-cb23","outbound_label":"DOI","outbound_source":"dc:identifier","record_url":"https://doi.org/doi:10.25335/zzrr-cb23","record_label":"Repository record","record_kind":"doi","record_source":"dc:identifier","download_url":null,"download_label":null,"download_source":null,"access_status":null,"embargo_until":null,"access_note":null},"source_record":{"url":"https://d.lib.msu.edu/oai?verb=GetRecord&metadataPrefix=oai_dc&identifier=oai%3Ad.lib.msu.edu%3Aetd_52748","prefix":"oai_dc"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Radha, Hayder","Morris, Daniel","Traganitis, Panagiotis","Liu, Xiaoming"]},{"key":"dc:creator","label":"Author","values":["Lu, Xiaohu"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2025"]},{"key":"dc:relation","label":"Dc Relation","values":["Electronic Theses & Dissertations"]},{"key":"dc:type","label":"Dc Type","values":["text","Electronic dissertations"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Engineering","Electrical engineering","Computer engineering"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["eng"]},{"key":"dc:rights","label":"Dc Rights","values":["In Copyright"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["etd:52748","local:Lu_grad.msu_0128D_21726","https://doi.org/doi:10.25335/zzrr-cb23"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Thesis (Ph.D.)--Michigan State University. Electrical and Computer Engineering - Doctor of Philosophy, 2025","Object detection using LiDAR point clouds relies on a large amount of human-annotated samples when training the underlying detectors' deep neural networks. However, generating 3D bounding box annotation for a large-scale dataset could be costly and time-consuming. Alternatively, Unsupervised Domain Adaptation (UDA) enables a given object detector to operate on a novel new data, with unlabeled training dataset, by transferring the knowledge learned from training labeled source domain data to the new unlabeled target domain. Pseudo-label strategies, which involve training the 3D object detector using target-domain predicted bounding boxes generated by a pre-trained model, are commonly used in UDA. However, these pseudo-labels often introduce noise, negatively impacting performance. Furthermore, current UDA 3D object detection methods tend to overlook the inherent geometric information in 3D point clouds, as well as valuable information from other modalities, such as images. In this dissertation, our objective is to address three key limitations and shortcomings of state-of-the-art unsupervised domain adaptation approaches for point-cloud based 3D object detection. We mitigate these limitations and shortcomings by developing three novel frameworks that: (a) Denoise pseudo labels generated by pre-trained models; (b) Extract and exploit 3D geometric information; and (c) Employ and inject 2D image information into 3D point clouds to improve LiDAR-based object detection in the target domain. Under our first contribution, we developed a new framework, DALI, which addresses two common forms of pseudo-label noise: distribution-level (global level) and instance-level (object level) types of distortions. In particular, DALI addresses the domain gab between different datasets that exhibit object sizes' variations, which results in target-domain noisy pseudo labels. In that context, DALI introduces a post-training size normalization strategy that reduces distribution-level noise by identifying an optimal unbiased scale after training. To mitigate instance-level noise, we build libraries of 3D models and LiDAR sensors and propose two pseudo point cloud generation methods — ray-constrained and constraint-free - that simulate LiDAR scanning to produce cleaner pseudo point clouds. To explore the benefits of geometric information in 3D point cloud, we developed our second contribution, ENGIN. Here, we propose an unsupervised geometric size estimation method that infers average object dimensions from accumulated point clouds of predicted boxes. We then design a geometric diffusion network to generate denoised, augmented point clouds for target-domain pseudo labels. ENGIN also introduces a BEV domain classifier with a domain embedding strategy to enhance joint training on source and target samples.To take advantage of valuable information from other modalities, such as images, in our third contribution, CMOS, we propose a framework for unsupervised domain adaptation in 3D object detection by leveraging cross-modality semantic information from images. The method jointly trains a 2D segmentor and a 3D detector using only 3D bounding box labels. Key components include bidirectional semantic injection (3D-to-2D and 2D-to-3D), semantic-consistent augmentation for aligned data transformations, and mutual-modality filtering to reduce false positives. We evaluated DALI, ENGIN, and CMOS on widely used public datasets, including KITTI, nuScenes, and Waymo. The superior experimental results demonstrate the effectiveness of the proposed frameworks.","Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).","Includes bibliographical references."]},{"key":"dc:format","label":"Dc Format","values":["application/pdf","1 online resource (94 p.) : ill."]},{"key":"dc:title","label":"Title","values":["DOMAIN ADAPTIVE 3D OBJECT DETECTION FOR AUTONOMOUS VEHICLES"]}]}],"canonical_facts":{"dc:contributor":["Radha, Hayder","Morris, Daniel","Traganitis, Panagiotis","Liu, Xiaoming"],"dc:creator":["Lu, Xiaohu"],"dc:date":["2025"],"dc:description":["Thesis (Ph.D.)--Michigan State University. Electrical and Computer Engineering - Doctor of Philosophy, 2025","Object detection using LiDAR point clouds relies on a large amount of human-annotated samples when training the underlying detectors' deep neural networks. However, generating 3D bounding box annotation for a large-scale dataset could be costly and time-consuming. Alternatively, Unsupervised Domain Adaptation (UDA) enables a given object detector to operate on a novel new data, with unlabeled training dataset, by transferring the knowledge learned from training labeled source domain data to the new unlabeled target domain. Pseudo-label strategies, which involve training the 3D object detector using target-domain predicted bounding boxes generated by a pre-trained model, are commonly used in UDA. However, these pseudo-labels often introduce noise, negatively impacting performance. Furthermore, current UDA 3D object detection methods tend to overlook the inherent geometric information in 3D point clouds, as well as valuable information from other modalities, such as images. In this dissertation, our objective is to address three key limitations and shortcomings of state-of-the-art unsupervised domain adaptation approaches for point-cloud based 3D object detection. We mitigate these limitations and shortcomings by developing three novel frameworks that: (a) Denoise pseudo labels generated by pre-trained models; (b) Extract and exploit 3D geometric information; and (c) Employ and inject 2D image information into 3D point clouds to improve LiDAR-based object detection in the target domain. Under our first contribution, we developed a new framework, DALI, which addresses two common forms of pseudo-label noise: distribution-level (global level) and instance-level (object level) types of distortions. In particular, DALI addresses the domain gab between different datasets that exhibit object sizes' variations, which results in target-domain noisy pseudo labels. In that context, DALI introduces a post-training size normalization strategy that reduces distribution-level noise by identifying an optimal unbiased scale after training. To mitigate instance-level noise, we build libraries of 3D models and LiDAR sensors and propose two pseudo point cloud generation methods — ray-constrained and constraint-free - that simulate LiDAR scanning to produce cleaner pseudo point clouds. To explore the benefits of geometric information in 3D point cloud, we developed our second contribution, ENGIN. Here, we propose an unsupervised geometric size estimation method that infers average object dimensions from accumulated point clouds of predicted boxes. We then design a geometric diffusion network to generate denoised, augmented point clouds for target-domain pseudo labels. ENGIN also introduces a BEV domain classifier with a domain embedding strategy to enhance joint training on source and target samples.To take advantage of valuable information from other modalities, such as images, in our third contribution, CMOS, we propose a framework for unsupervised domain adaptation in 3D object detection by leveraging cross-modality semantic information from images. The method jointly trains a 2D segmentor and a 3D detector using only 3D bounding box labels. Key components include bidirectional semantic injection (3D-to-2D and 2D-to-3D), semantic-consistent augmentation for aligned data transformations, and mutual-modality filtering to reduce false positives. We evaluated DALI, ENGIN, and CMOS on widely used public datasets, including KITTI, nuScenes, and Waymo. The superior experimental results demonstrate the effectiveness of the proposed frameworks.","Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).","Includes bibliographical references."],"dc:format":["application/pdf","1 online resource (94 p.) : ill."],"dc:identifier":["etd:52748","local:Lu_grad.msu_0128D_21726","https://doi.org/doi:10.25335/zzrr-cb23"],"dc:language":["eng"],"dc:relation":["Electronic Theses & Dissertations"],"dc:rights":["In Copyright"],"dc:subject":["Engineering","Electrical engineering","Computer engineering"],"dc:title":["DOMAIN ADAPTIVE 3D OBJECT DETECTION FOR AUTONOMOUS VEHICLES"],"dc:type":["text","Electronic dissertations"]},"updated_at":"2026-11-18T18:59:26Z"}