TOKYO — NEC Corporation has announced the development of a “gradual deep learning-based object detection technology” that enables efficient and high throughput object detection for video analytics while maintaining detection accuracy.
This technology enables up to eight times the processing speed of object detection for large volumes of images, even on edge devices with limited processing capacity.
NEC aims to commercialize this technology in fiscal 2022, following further research and development.
Video analytics are expected to be utilized for a wide range of applications, such as analyzing camera images of vehicles at intersections, optimizing traffic control, and analyzing camera images of stores and warehouses to detect intrusion or to optimize facility management.
To perform these video analytics in real time, it is ideal to process them on an edge device near a sensor, such as a camera.
However, because cooling is difficult to manage and electricity consumption is restricted in edge devices, high-performance processors such as GPUs used in high-performance servers are not available, and processing capacity is constrained.
In video analytics, object detection software that utilizes deep learning (hereinafter “object detection AI model”) performs object detection processing in order to find the object to be analyzed from images captured by a camera
Since highly accurate object detection AI models have a large amount of operations though, it is difficult for edge devices to process a large amount of images due to constraints on processing capacity. If the amount of operations for a high-speed object detection AI model is reduced, for example, the accuracy declines and the recognition accuracy requirements for image analysis cannot be met.
Application of NEC’s newly developed gradual deep learning-based object detection technology enables efficient, high-speed, and high-precision detection of subjects from a large amount of images, even in an edge device with limited processing capacity, and enables simultaneous processing of images from multiple cameras in real time.
1. High-speed, high-precision detection at the same time
This technology combines an object detection AI model that is high-speed, but is only roughly accurate, with another object detection AI model that is highly accurate, but computationally complex. This combination enables efficient processing of multiple images and results in high-speed, high-precision detection.
First, a high-speed object detection AI model detects multiple objects quickly, but with rough accuracy, then the detected results are processed together with a high-precision object detection AI model to gradually refine the detected objects.
As a result, in an example of detecting a car license plate from a camera image, processing speed was approximately eight times faster than when using an object detection AI model with high accuracy alone, while maintaining detection accuracy*.
2. Compatible with various detection targets, detection methods, and AI chips
This technology is compatible with a diverse range of object detection processes according to detection targets, such as people, vehicles and car license plates, without being limited to specific detection targets and specific processing methods.
In addition, the technology’s application is not limited to a specific AI chip. By taking advantage of the fact that AI chips have a large number of computing resources in common, the technology makes effective use of the internal computing power of chips.
Going forward, NEC aims to enhance the safety, security, and convenience of society by expanding the use of video analytics technologies to a wide variety of applications.
*Speed varies depending on the device used, the type of detection engine and the quality of image data.