Description of original award (Fiscal Year 2020, $836,000)
This study proposes (1) to develop novel multi-modal techniques to automate analysis of audio and video data from body-worn camera (BWC) recordings of police and community interactions to evaluate officers adherence to principles of procedural justice and (2) to validate the ratings generated by the automated process using an randomized control trial (RCT) comparing software ratings of BWC videos to evaluations performed by human raters under conditions of high and low procedural justice.
The multi-modal analysis proposed here take both audio and video data from BWC footage and integrates the two into representations of the interactions between officers and members of the public using cutting-edge Natural Language Processing and Computer Vision tools. These representations take simple features of language and behaviors such as body movements, facial expressions, and verbal utterances, and other features of interactions captured by BWCs, and combines them to generate more complex understandings of phenomenon related to procedural justice such as rapport, respect, trust, cooperation, conflict, coordination, and so forth. Multi-modal analysis enables researchers to go beyond the static measurement of individual traits and states (e.g. emotion, personality, etc.) and explore the dynamic relationships between individuals that unfold in real time during every human encounter. By applying rigorous computational models of social interaction to the relationships captured in BWC footage, the project team aims to gain unprecedented capacity to open a window onto the actual procedures of language and behavior that underlie the academic concepts of procedural justice. During the development phase of the project, Polis Solutions and GE Global will analyze a dataset of 100 hours of BWC footage of contacts between officers in the Dallas Police Department and pedestrians in a wide range of settings as ground truth. Next, the proof of concept phase will involve the random assignment of police officers in the treatment group receiving procedural justice training or the control group that will not receive the training. Each of the 20 officers will be involved in 40 interactions with citizens captured on video (total N=800 incidents).Using mixed effects regression models, the research team will test: (1) whether videos of officers who received the training showed greater adherence to procedural justice principles and (2) whether there were no significant differences between video ratings by the automated tools and manual raters among either the 400 videos from the low procedural justice (untrained) officers and the 400 videos from the high procedural justice (trained) officers. Letters of support from all of the participating organizations provided in the application, including the Dallas Police Department.
Note: This project contains a research and/or development component, as defined in applicable law, and complies with Part 200 Uniform Requirements - 2 CFR 200.210(a)(14). CA/NCF