After almost a decade of work, researchers at UCSD created an algorithm that can detect pedestrians in near real-time. The algorithm is able to capture images at a rate of about two to four frames per second and achieves a higher level of accuracy than any existing algorithms.
UCSD professor and lead researcher Nuno Vasconcelos told the UCSD Guardian that his goal is to create an algorithm that allows computers to process scenes as well as humans.
“Currently, the solution only applies to the detection of a single class of objects, like pedestrians,” Vasconcelos said. “We are working on extensions that will detect multiple object-classes (say pedestrians and vehicles) simultaneously.”
According to a Feb. 8 press release, the algorithm achieves its speed and accuracy by breaking up the object detection process into several stages, where each stage simplifies the image. The first few steps remove parts of the image that obviously do not contain people, such as the sky. In each subsequent step, the computer studies the remaining picture more closely and removes some of the objects that do may resemble people, but are not pedestrians, such as trees, until only pedestrians remain. This method of object detection is called cascade classification.
Vasconcelos and his colleagues created many versions of the cascaded classifiers algorithm before reaching the current design, which is faster and more accurate than its predecessors.
“The research on cascaded classifiers started around 2008 [when] deep learning was not around, but the project allowed us to learn to design these cascaded classifiers,” Vasconcelos explained. “The main difficulty [is finding] the optimal cascade configuration, which we addressed by designing this new cascade learning algorithm [using deep learning].”
The speed and accuracy of the new algorithm makes it great for use in cars to identify pedestrians and other objects. Zhaowei Cai, a Ph.D student who has been working with Vasconcelos on this project since 2014, explained that while the algorithm is not ideal for self-driving cars, it can still be used to assist drivers with parking and braking.
“We work with some companies [that] want to build the detector into cars, maybe just to assist the driving, not just for self-driving,” Cai told the Guardian. “It can help to detect people or objects nearby.”
Although the research is currently being used in cars, Vasconcelos added that the technology can also be applied to other areas, such as medicine and retail.
“Object detection and recognition are central problems in computer vision,” Vasconcelos explained. “The technology could, for example, be used to detect tumors or lesions in medical images, products in product catalogues, or most other applications that involve image or video understanding.”