The faces of about 7,000 people have been loaded into a database for Edmonton Police Service (EPS) and a body-worn camera manufacturer to test facial-recognition software in the field. Starting Wednesday, for the duration of December, the facial-recognition software will run on footage captured by EPS members conducting investigative or enforcement work, searching for matches in the pre-loaded database. All facial data captured within four metres of a police officer’s body camera will be sent to a cloud for comparison by the artificial intelligence (AI) program; if it does not match, the data will be discarded. Potential matches, however, will be sent for review and confirmation by a trained professional within about 24 hours. “If our officers executed three warrants, and there was the potential that they could have executed 50 warrants and we missed all those warrants, that – for me – would basically highlight the potential value of this technology,” said Supt. Kurt Martin of EPS’ information and analytics division, during a news conference on Tuesday. The pilot will also be a test of feasibility in Edmonton’s winter climate, as low light can affect the ability of the facial-recognition software. If implemented in the future as standard practice in the field, officers would be notified in “near real time” of confirmed identity matches with safety risks or outstanding warrants, Martin said. “At this stage here, we’re just evaluating it to see if there’s potential applications, but none of the results of this first proof of concept will change police behavior. They’re not even going to be notified about the potential resemblance notifications at all,” he told reporters. How the software works EPS’ pilot is the first time the facial-recognition software will be trialled, and as such is a test for the service’s body camera manufacturer Axon Enterprise, too. The pilot will have a minimal cost for EPS. Axon cannot view or control the facial data that EPS enters in its database, which is stored on a secure server at an undisclosed location in Canada. The 7,000 faces consist mostly of people who have been flagged as a safety risk, as well as a few hundred who are wanted on at least one serious warrant, such as murder, aggravated assault and robbery. Although EPS’ body-worn cameras are “passively” recording whenever they are worn, the facial-recognition software will not kick in until the officer switches to “active” recording mode for investigative or enforcement purposes, according to Martin. To demonstrate, Martin took a photo of Axon’s director of responsible AI, Ann-Li Cooke, standing amidst other spokespeople and reporters during the Tuesday press conference. The program only identified her face, matching it to a picture that had previously been uploaded to the system. The software falls on the machine-learning model side of the artificial intelligence spectrum, meaning it has been trained with existing data to make its decisions. “This is, in my mind, something that’s very similar to the Automated Fingerprint Identification System that we’ve been using for, like, 30 years. It just takes it to a different level. We’re now using faces instead of fingerprints,” Martin said. EPS also uses similar technology to identify faces in security footage. Martin said using it in the field has the potential to make the work of officers safer and faster. Today’s limitations of facial-recognition software Cooke declined to say which third party’s machine learning model Axon was using for the pilot before a contract had been finalized. Axon had previously abstained from riding the facial-recognition software wave for concerns about demographic bias and overall accuracy. “All of the models that have the highest ratings have done a lot of research and a lot of training of their models to fill in the gaps that were previously there in 2019,” Cooke told reporters on Tuesday. Since then, the industry has eliminated race-based differences in ideal conditions, according to Cooke. She noted, “Today, the limiting factor is on skin tone, and so when there is very varying conditions, such as distance, dim lighting, there will be different optical challenges with body-worn camera – all cameras – in detecting and matching darker-skinned individuals than lighter-skinned individuals, and that’s why we’ve put in guardrails in this technology to limit the distance and also to be only used in good lighting, until we learn more about the capabilities.” She declined to quantify the software’s error rate. Privacy commissioner’s assessment EPS says it has submitted a privacy impact assessment to Alberta’s information and privacy commissioner to ensure the pilot is fair and lawful. Typically, the commissioner’s recommendations from such assessments are not made public but commissioner Diane McLeod promised her team would be examining Axon’s technology closely. She has concerns about EPS fulfilling its duty of accuracy under the Protection of Privacy Act, as well as demographic biases the facial-recognition industry has historically grappled with. “The impacts to the public are significant if we have mismatches: false positives, false negatives. This is policing. This is serious. This has real impact on people’s lives,” McLeod told CTV News Edmonton. With files from CTV News Edmonton’s Gates Guarin
