Computer Vision News - June 2022

19 Liberty Defense Dan adds: “ To some degree, it restricts the types of algorithms we can apply . We need something that can finish – if we’re doing 30 hertz, you’ve got about 30 milliseconds to be done, and that’s the whole pipeline. Unlike a video camera where you get an image already formed, we have a processing step called holographic reconstruction. If you’re familiar with CT, it’s like tomographic r e c o n s t r u c t i o n , but it’s essentially reconstructing from the various wavefronts captured by the sensor. ” T r a d i t i o n a l l y , reconstruction comes first, then AI, but here, combining the reconstruction process end to end with the AI process into one optimization was the only way to achieve the speed and performance enhancements required. The team looked to the autonomous driving community for guidance, where several single-shot detector algorithms have been developed , sacrificing some detection performance but being designed to be very fast. “ That ends up being a big win for us, ” Dan says. to pass through with permissible objects , such as keys or a cell phone, without causing false alarms and without needing them to remove clothing items every time. That’s where AI is going to be key. ” Security cannot be compromised to achieve this high throughput; it should be enhanced over current products. The process should be almost invisible to people, non-intrusive, andprotecttheirprivacy. To be successful and deployable, it must use a low-cost computational platform, as venues like schools and stadiums do not have the money to pay for costly solutions. The team’s new system uses millimeter-wave imaging , which can penetrate clothing but not skin. It supplies a complete data set at speed, with the AI making swift decisions. “ We’re creating whole datasets, with hundreds of thousands of data points available for reconstruction at 10 to 20 times a second – eventually 100 times a second , ” Jeff tells us. “ Think of a fast-frame camera. When someone walks by, it’s as if they’re frozen, but they’re moving. You must go through image processing, reconstruction, and AI decision-making 20 times a second. It’s a pipeline that’s very demanding. ” Dan Pineo Jeff Gordon