No Cameras Please

Object detection technology has been making significant strides in recent years, and it is increasingly being used across various industries to improve efficiency, productivity, and safety. This technology is most commonly based on computer vision algorithms that enable computers to identify and locate objects within digital images or videos.

The applications of object detection are numerous and varied. In the retail sector, it is used to track inventory levels, monitor customer behavior, and detect shoplifting. In the transportation industry, it is used for traffic monitoring, license plate recognition, and autonomous driving. In healthcare, it is used to detect cancer cells, track the movement of organs, and aid in medical diagnosis.

Computer vision-based object detection algorithms can require a significant amount of computational resources and energy, however. This is because these algorithms analyze individual pixels and features within images or videos to identify and locate objects, which requires a very large number of calculations. And as the amount of data to be analyzed increases, so does the computational demand, making it challenging to process high resolution scenes in real-time.

A promising new technique has recently been described by a team of researchers at the Beijing Institute of Technology that has the potential to perform real-time object detection without the need for a lot of computational horsepower. Their image-free method uses a single pixel detector and a machine learning analysis pipeline to detect objects with fairly remarkable levels of accuracy. And whereas previous image-free efforts have failed to get the class, location, and size information of all objects in a frame at the same time, this team has shown that their system is up to the task.

This novel single-pixel object detection technique illuminates an area with a carefully crafted sequence of structured light patterns, and the intensity of the light is recorded by a single-pixel detector. This scanning process is very quick, requiring just a tiny fraction of a second to complete. And because the measurements are sparse, the algorithms that process them can be lightweight.

The data was first processed by a transformer-based encoder, which was used to extract the most relevant and informative features from it. These features were then forwarded into a multi-scale attention network-based decoder that is capable of predicting the class, location and size of all detected targets at the same time.

By engineering an optimized structured light pattern for the system to emit, and leveraging a multi-scale attention network, the algorithm was able to focus on the important aspects of a scene. This, in turn, enabled the technique to efficiently extract features from a scene, and to deliver state-of-the art object detection performance.

To assess the performance of the method, it was tested on images from the PASCAL Visual Object Classes dataset. Images from the dataset were printed on film and scanned by the system, and an average detection accuracy of 82.2% was observed across all object types represented in the dataset. These detections were fast as well, clocking in at about 63 frames per second. The number of detectable objects is presently limited to the 80 that appear in the Visual Object Classes dataset, but with additional training, this new method could, in theory, detect any other types of objects as well.

The unique characteristics of this approach make it well-suited for tinyML and edge computing applications. Perhaps we will see single-pixel object detection powering the algorithms for autonomous driving and more in the future.