Page 14 - Security Today, September/October 2023
P. 14
Scene Intelligence Evolution How video analytics are becoming smarter and more accurate By Robert Muehlbauer Remember the early days of video analytics? All the search algorithm to successfully locate a certain object of a pre- cise color, it needs to be taught to recognize color and how to dis- alerts triggered by a passing shadow? Or leaves tinguish the differences between colors. If an algorithm is meant quivering in the breeze? Even a car with bright headlights driving by? Because these analytics to trigger an alert on a particular type of vehicle, the analytics have to be programmed to discern the differences between motor- were based solely on pixel changes, they tended to cycles and bicycles, trucks and cars, buses and mobility scooters. generate a lot of false alarms. In some cases, the number of false alarms generated by these first analytics became so frustratingly With further training, a model could even learn to recognize and classify specific vehicle models from a chosen manufacturer. high that some users decided to simply turn them off altogether. Fast-forward and today you will find that video analytics has Depending on the complexity of the application and the num- come a long way thanks to better image processing and deep-learn- ber of variables the algorithm needs to classify, developers might need to rely on big data sets to train their deep-learning modules. ing software models trained to discern differences between objects and people. This makes it possible for the camera to capture high- The amount of data sets needed to support the analytics gener- ly-granular metadata – such as the color of clothing, the type of ally determines where the analytics should reside. If the data set is relatively small – such as detecting whether a person is loitering or vehicle, the direction an object is traveling – which makes it easier to locate and track the movement of people and objects through crossing into a restricted zone – the analytic can reside in-camera. a scene, whether in real-time or when searching archived footage. Placing analytics at the edge reduces latency and delivers greater accuracy since the video does not need to be compressed PROVIDING A FOUNDATION FOR DEEP LEARNING – and thus possibly lose critical details – when being transmitted How do developers train these more advanced analytics? They to a server for analysis. But a camera’s system chip and deep- are taught by example. The more examples they are given for learning processor need to be sufficiently robust for the task. comparison, the more accurate they become. For instance, for a For applications requiring larger data sets – like reading and ART STOCK CREATIVE/Shutterstock.com 14 SEPTEMBER/OCTOBER 2023 | SECURITY TODAY INTELLIGENT VIDEO