Video evidence, from dash and body-worn cameras, surveillance systems and smartphones, has given rise to a new dimension of police investigations. There is often so much video evidence that the investigator is overwhelmed by the task of culling out the irrelevant segments to focus on those that provide genuine insight to the incident. Video analytics powered by artificial intelligence (AI) simplifies and accelerates that process.
Investigators may be forced to watch many hours of uneventful video, looking for the brief segment that contains evidence. A similar problem is when someone is assigned to monitor multiple live feeds of surveillance video, and is to send out an alert when something noteworthy is observed. In both cases, the viewer’s mind can start to wander, and it’s not long before the person watching the video is shown exactly what he or she is looking for, and misses it completely. Research has shown the people who watch a single video monitor for more than 20 minutes lose 95% of their ability to detect events displayed on the monitor. Humans get bored, but the attention span of a computer is infinite.
One of the simpler AI applications in video analysis involves the designation of a zone in the video frame, where any movement causes an alert to generate. Retail loss prevention officers use this feature to call attention to activity around high-value merchandise, or to review old footage to determine when an object was removed from a shelf.
The analyst draws a polygon around the area or areas to be monitored. If the pixels within the boundary change, that portion of the video is flagged for review. With a real-time alert, the operator is notified of any change within the boundary, so they can watch to see if the movement is something innocuous, or represents a hazard that merits a human response.
In reviewing old footage, the zone alert relieves the operator from having to watch hours of what is essentially a static scene. Clips that show activity within the designated zone are marked automatically, and the operator can look at each one to determine if there is evidence there.
Motion or absence of motion detection
A bag or parcel left behind in a crowded area might be simple forgetfulness, or the start of a terrorist attack. The presence of an abandoned or otherwise suspicious package in an area with high foot traffic can signal the presence of an improvised explosive device. The AI in a surveillance system can watch for objects that are not part of the standard background that don’t move within a few seconds. It will ignore a briefcase that is placed on the ground while a commuter checks his smartphone, but zoom in on one that is left behind when the owner walks away. Once the suspicious package is identified, officers can be dispatched to investigate.
The velocity of a moving object is also a potential trigger for the system. People who are on the run merit greater attention. Runners create a hazard, but more importantly, they may have just assaulted someone or stolen a purse, and are making their getaway. This rapid movement can generate an alert to get police to the area, or at least provide a direction of travel for the suspect.
Artificial intelligence can be “trained” to recognize large moving shapes as vehicles, and that vehicles should be moving in one direction within a video boundary, or maybe not be present at all. This type of programming can alert the operator if a vehicle is proceeding down a street in the wrong direction, or encroaches onto a sidewalk.
Location of objects
More advanced video AI can locate objects in a video frame after being given description parameters or an example. In a simpler application, AI can search a large volume of video for the presence of a certain model of vehicle, once the dimension, shape and even color of that vehicle is provided to it. AI now available will identify every moving object in the video frame, and find those same objects in video from other sources. This technology was put to use in the investigation of the 2013 Boston Marathon bombing, when investigators were called to analyze many hours of video from dissimilar sources, trying to find the appearance of the suspected bombers or their vehicles.
Another application of AI in video analysis is redaction of details from video, when privacy laws or investigational integrity necessitates blurring or blocking elements shown in the video. These elements range from mailboxes and house numbers to license plates and faces of bystanders.
Manual redaction may require drawing boxes around the objects to be redacted in every frame of the video, a process that is both tedious and time-consuming. Automated redaction systems allow the operator to identify the object to be redacted, and then let AI track that object across multiple frames, blurring or blocking the display of the object.
The ability of artificial intelligence to recognize the faces of persons of special interest is controversial. This technology has improved markedly with the availability of higher-resolution cameras, faster transmission of data streams, and greater computer processing power. Yet, some privacy advocates object to the “surveillance society,” where the movement and activities of everyone can be or are monitored by the government or employers.
From a law enforcement perspective, this technology is gold. Fixed surveillance cameras can alert monitors to the presence of wanted persons. Police officers with body-worn cameras may soon be notified of the identity of people they encounter in the field via a head-worn display or other alerting system, and know when someone has provided them with a false name.
One maxim of computer science is that “computers work; people should think.” The increased use of AI in video analysis allows law enforcement to do just that. Relieved from some of the more tedious aspects of evidence review, investigators can be more efficient and better serve their communities.