As IP cameras become cheaper and easier to activate, security cameras are now pervasive in homes and businesses. However, no human can digest dozens of video streams continuously. We get tired, distracted and easily lose focus on monotonous activity.
Internet of things (IoT) sensor-based technologies that detect movement or other environmental variables work well in testing, but they trigger too many false positive alerts in practice to be useful in most cases. As a result, billions of hours of video footage is often used only for post-incident investigation, instead of catching crime when it is happening.
Is there a better way? Is it possible to provide real-time alerts to prevent crime before or while it is occurring, without inundating users with false positives or false negatives?
The latest computer vision techniques may provide some answers.
In this article, we look at how artificial intelligence (AI)-driven computer vision harvests value in real-time from security video data feeds, extracting actionable insights while preserving privacy, with minimal human involvement.
Computer vision is considered a subfield of artificial intelligence and machine learning. It focuses on helping computers see and understand the content of digital images. Recent shifts from statistical methods to deep learning neural network methods help computer vision in developing a variety of security uses. Here are four key ways advances in computer vision are improving physical security:
1. Recognizing and labeling an object/pattern
Recognizing objects and patterns has many applications: automatic medical diagnosis (health care), defect reduction (manufacturing), pest infestation prediction (agriculture).
In security, computer vision does even better than human eyes in detecting humans, vehicles or guns. Sentry AI, a Silicon Valley startup that I had invested in, partners with major video management software and camera makers to empower AI on any camera. For little cost, a regular camera can accurately detect a human presence from different angles even when only part of an arm appears in the field of view.
2. Answering the age-old question: Who are you?
Authentication through facial recognition entered the mainstream through iPhone’s FaceID. The implementation of facial recognition, however, preserved privacy by storing data only on the user’s device. Lately, it’s being used for many other use cases where biometrics are stored in central databases. (See the Hertz car rental partnership with Clear as an example.)
There are general-purpose facial-recognition solutions offered by public cloud vendors such as Amazon, Google and Microsoft. Computer vision scientists are also developing other approaches to identify people based on their geometric shapes (height, width and body-part proportions) and gait cues (stride length and amount of arm swing). Combining these approaches promises higher accuracy on person identification under different camera angles, poor light conditions and long distances.
3. Inferring actions from a sequence of images or video
Computer vision can help digitize specific events, times and locations, and can use this data to track behavior. For example, a camera in a retail store can track employee activity in real-time, alert when a new customer enters or exits the store, and track their journey. This information can not only be used to detect loitering outside the stores and reduce shoplifting, it can provide actionable insights for improving traffic flow and placing merchandise.
4. Modifying or recreating realistic images
A computer vision technique called "generative adversarial networks" (GAN) can be used to generate photo-realistic images, reconstruct damaged images and remove blurring and partial obscuration from rain. (Look here to see how this technique could be used in practice.) This technique would be useful in generating visualization for critical security incidents and reconstructing faces or license plates to provide law enforcement richer information.
With these new techniques and rapidly improving capabilities, computer vision is progressing toward solving certain security challenges. Scientists all over the world are making rapid advances, and companies are making AI useable by providing platforms that could be used by nonspecialists.
Google’s Tensorflow is an open-source AI-software platform that anyone can use. Labeled datasets, needed for training and testing, are becoming widely available. (Take a look at the labeled data platform from Scale as an example.)
AI could watch cameras tirelessly, monitor people and patterns and look for indicators of security concern. AI could systematically search a field of view for objects of interest or correlate movements and look for anomalies. Privacy is enhanced as AI processes data in a black box and can remove the images/records after inference.
Just like any other powerful technology, computer vision has both benefits and serious concerns. The rise of AI in surveillance stirs fears about loss of privacy and government intrusion. Under the guise of security and crime prevention, AI could be abused to track people and control behavior. In China, for example, SenseTime has become one of the largest providers of AI in the security space. Its technology is being used for everything from authenticating people by recognizing their faces to detecting crime as it happens. Such a technology would raise serious privacy concerns outside of China.
Moving forward, we need regulations to enforce disclosure of how companies gather and use the data. Such policies should cover not just the captured images but also the information inferred from those images and how that information is correlated with other data sources to drive conclusions.
Conclusion
By better deriving intent from objects and motions and filtering out the false alarms, computer vision shows enormous potential as a risk reduction tool and an information filter to help security personnel be more effective at their jobs. With more active researchers in the field, we expect to see far more accurate and reliable vision in the future. The day AI can guard you while you sleep may not be that far off.
