Researchers Tout Advanced Pedestrian Detection System

Researchers at the University of California-San Diego (UCSD) say they have developed a pedestrian detection algorithm that is faster and more accurate than current options.

The new system combines aspects of “cascade” detection techniques with so-called deep-learning technology. This allows for near real-time vision while cutting error rates in half, which the researchers note will be critical for self-driving vehicle features.

As with current approaches, the new algorithm simultaneously analyzes millions of individual pieces of a video image to determine the presence of a pedestrian. This is done in three basic stages. Segments that clearly contain no visible pedestrians are quickly assessed and eliminated, followed by those showing some human-like characteristics (a similar sized or shaped object). In the more complex segments, the software must distinguish between a pedestrian and very similar objects.

Whereas current identification technologies use a cascade detection for all three stages with an increasing number of classifiers—all of which are relatively simple predictors—at each stage, the UCSD program adds deep-learning models for the more complex image pieces. The latter allows for advanced pattern recognition learned from thousands of similar examples in which the presence or absence of a pedestrian has been determined previously.

Currently, the new technique can be used only to detect one pedestrian at a time. The next step of the research will focus on detecting multiple pedestrians.

The UCSD team is led by Nuno Vasconcelos, an electrical engineering professor at the university’s Jacobs School of Engineering. Vasconcelos also is a faculty advisor for the Center for Visual Computing and the Contextual Robotics Institute, both of which are housed on UCSD’s campus. The research was supported by awards from the National Science Foundation and funding from Northrop Grumman.