AI analysis of body-worn camera footage to study and improve traffic stops — NSF AI grant $1.25M (S&CC)

The NSF awarded about $1,250,000 to develop AI tools that analyze body-worn camera footage to study and improve traffic stops (NSF Award 2531357; program: Smart & Connected Communities; California).

1Traffic stops as complex interactions

Per the abstract, police traffic stops are common, complex interactions that can be a tool to improve public safety or escalate to violence. Effective communication is crucial for officer and civilian safety, enforcing the law, and building public trust.

The project develops AI tools enabling researchers to analyze footage from officers' body-worn cameras, learn about officer–driver communication, and refine best practices for traffic-stop outcomes.

2A partnership with LAPD and community groups

To do so, it draws on collaborative research capacity built among academic researchers, the Los Angeles Police Department, and over a dozen community organizations. The multidisciplinary, community-engaged approach ensures the AI tools reflect a wide range of viewpoints and address stakeholders' concerns.

The tools are trained on assessments created by individuals from varied backgrounds, including retired police officers and Angelenos with a mix of past positive and negative experiences with police. After development, the tools can be used by police departments and local governments nationwide to lower costs and enhance transparency, accountability, and learning.

The project advances computer, social, and engineering science by developing video language models that incorporate multiple stakeholder perspectives and infrastructure for collaborative AI development.

3From footage to a working tool

The project sets out to turn footage that already exists into something usable in the field, and the path runs in stages.

1. 1Gather footageWork with the Los Angeles Police Department on records of traffic stops
2. 2Attach ratingsHave retired officers and residents of varied backgrounds rate the footage
3. 3Build the modelTrain video language models that take in several perspectives
4. 4Return it to the fieldPut it in a form police departments and municipalities nationwide can use

What sits at the centre is not building the technology first but bringing people of different standing into the stage where ratings are attached. Who decides what counts as good changes the character of the tool that results.