A team of researchers from the University of California, San Diego (UCSD) have developed a pedestrian detection system that, according to the developers, performs in near real-time and with higher accuracy compared to existing systems. 

The research was directed by Nuno Vasconcelos, electrical engineering professor at the UC San Diego Jacobs School of Engineeringwho stated that their biggest goal is real-time vision (2–4 frames per second), especially for pedestrian detection systems in self-driving cars, but this algorithm and technology could also be used in robotics, and other image and video search systems.

Vasconcelos and his team combined a traditional computer vision classification architecture, known as cascade detection, with deep learning models in order to improve accuracy and speed of pedestrian detection systems.

The pedestrian systems usually break down an image into sections known as small windows, which are processed to determine if there is a pedestrian present or not. This method poses quite a challenge as it is difficult for the system to recognize pedestrians of different shapes and sizes (appearing near or far away from the camera) in various locations on the same image. Typically, it takes millions of windows inspected by video frame at speeds ranging from 5–30 frames per second.

The cascade detection technique that is used in the UCSD system basically does the exact same procedure, but instead of inspecting all the windows at once, it does the job in a series of consecutive stages with a new algorithm that incorporates deep learning models (better suited for complex pattern recognition) in the final stages of a cascaded detector. 

At first, the algorithm detects and discards all windows not containing a person; after that, it processes windows that require more 'work' from the system, namely classifying objects that might look like a person (a tree, fro example). In the final stages, the algorithm focuses on telling apart pedestrians from very similar objects.

"No previous algorithms have been capable of optimizing the trade-off between detection accuracy and speed for cascades with stages of such different complexities. In fact, these are the first cascades to include stages of deep learning. The results we're obtaining with this new algorithm are substantially better for real-time, accurate pedestrian detection," said Vasconcelos.

Feb. 12, 2016 Living photo: University of California, San Diego

This website uses cookies.
To comply with the EU regulations you must confirm your consent to their use.

You can do that by clicking "OK" or simply continuing to browse this website.
If you do not wish to have cookies set, you can opt out in cookie settings