There are wide and varied applications for machine learning; however, few are as compelling as those in public safety. Recently, I worked with the team from Arcturus who have been working on methods to provide operational insights and improve public safety for smart public transportation systems during NXP Connects in Santa Clara.
The application that Arcturus focuses on processes video from a security camera on a subway platform and looks for three conditions;
If one of these conditions are met, the system immediately provides a notification or performs a local, real-time action –such as shutting down track power. The real-time response really illustrates the power that machine learning has to transform a task that would typically be left to human observation to turn it into active detection.
To build the system, Arcturus chose NXP’s quad core i.MX 8M Mini applications processor combined with ArmNN to run their neural network detection algorithm at the edge. For a public safety system like this one, edge processing improves response time and reliability by eliminating the need to ship pixel data from each camera across the network and up to the cloud.
The table below describes the efficiency Arcturus was able to achieve using Arm NN, when compared to their existing pipeline deployed with OpenCV – the performance improvement allowed them to run their detection natively on the four Arm® A53 cores in the i.MX8M Mini, without the need for specialized ML hardware or even a GPU.
To help Arcturus port their application from OpenCV to the i.MX 8M Mini, we let them loose on a pre-production version of our eIQ™ machine learning software development tool. It’s important to note that eIQ also supports OpenCV for machine learning, but we’ll typically recommend using Arm NN for performance and efficiency reasons. Using eIQ software helped them rapidly deploy their application in days, rather than weeks without any direct support from us. From our perspective, this project offered a good exercise to ensure that eIQ machine learning development environment could deliver on the promise of a smooth transition from desktop-to-embedded application and that it offers users full advantage of the processing available in the i.MX 8M Mini device.
We’re excited to see what Arcturus has forthcoming. They have work underway to apply ML techniques that use clothing characteristics and reidentification methods to actively locate people. Therefore, losing your child at the mall may just become a thing of the past! If you need more information, contact them at www.ArcturusNetworks.com.