Innovation Lab

Public safety will continue to implement AI services with little understanding of the data that the AI models were trained on and insufficient capability to evaluate the effectiveness of the models. While this is an important topic in AI research, there is little to no research being conducted for public safety stakeholders.  n addition, public safety is reacting to current use cases and not planning for future AI use cases.​ 

• Document characteristics of quality AI datasets and the curation process and use this knowledge to identify potential AI resources needed to support PSCR research efforts.​  

• Document examples of the most impactful AI use cases and associated datasets to facilitate feedback from stakeholders on PSCR priority focus areas. ​  

• Design and host a stakeholder event to collect feedback on potential public safety AI use cases and select the most feasible and impactful datasets.​ 

• Document the stakeholder event report to inform FY27 project planning.​  

• Educate public safety stakeholders on how identify and prepare clean machine learning datasets.  

• Collect and present feedback on the importance of clean machine learning datasets and provide examples of datasets to support stakeholder validation. 

Reality capture can rapidly create 3D maps to improve first responder safety and situational awareness, but this geometric data lacks essential context. AI can provide this meaning, but its effectiveness is limited by a critical bottleneck: the scarcity of 3D data labeled with public safety-relevant information. This prevents AI models from being trained to perform reliably in new and unpredictable environments, which is a crucial requirement for emergency response tasks. 

• Develop an open-source software framework to study, measure, and compare self-supervised and multimodal learning approaches for interpreting public safety-relevant point cloud data. 

• Train an initial suite of machine learning models. 

• Execute model inference on representative test datasets. 

• Use customized analysis tools to objectively quantify model performance, generalization, and robustness. 

• Release the complete software framework, datasets, pretrained models, and generated labels to enable repeatable evaluation, benchmarking, and deployment by the research and public safety communities. 

• Enable scalable development of 3D scene understanding models. 

• Provide objective performance data that informs how well AI models interpret indoor environments for public safety operations. 

• Deliver open, validated tools and pretrained models that accelerate adoption and adaptation of AI-enabled reality capture. 

• Establish reproducible methods and benchmarks that support future standards, evaluation frameworks, and best practices.