Satya Swaroop Dash
Satya Swaroop Dash
Computer Science
09 Feb 2016

New Algorithm Helps Autonomous Cars Detect Pedestrians More Accurately: UC Research

Autonomous cars are here to stay, so why not make them as responsive as humans when it comes to detecting obstacles on the road such as pedestrians? UC San Diego researchers have devised a pedestrian detection algorithm that performs in near real-time with higher accuracy than existing systems. The algorithm developed by Nuno Vasconcelos, electrical engineering professor at the UC San Diego Jacobs School of Engineering and his team combines conventional computer vision architecture known as cascade detection along with deep learning models.

Pedestrian Detection UCSD

Normally, pedestrian detection systems break down the image captured by on-board cameras into small windows. These windows are then processed by a classifier to check the presence of pedestrians. Vasconcelos points out that this system requires much more processing power as pedestrians appear in different sizes (owing to their distance from the camera) and millions of windows have to be inspected in a video running at 5 to 30 frames per second. During the first stage of cascade detection, the detector separates the windows which do not contain pedestrians such as empty road or sky. In the second stage, it processes objects such as trees and in the final stage it distinguishes pedestrians with similar looking objects. The complexity is reduced in the final stage as the system has to process very few windows instead of millions.


The algorithm developed by Vasconcelos and his team assigns simple classifiers (weak learners) in the early stages of cascade detection and applies complex classifiers (deep learning models) in the later stages. This combination helps in much faster detection of pedestrians by achieving optimal trade-off between detection accuracy and complexity for each cascade stage, a feat that has not been achieved before. The only obstacle that prevents UC San Diego researchers from putting their creation into the real world is that the current system can only identify one object at a time. A smart car needs to detect a dozen or more types of obstacles on the road. The team is planning to have the algorithm to detect multiple objects to make it more practical. Once they perfect that you will be able to find the use of this algorithm in not only smart cars but in robots, security cameras and more.

Source: UCSD via Phys.org

Be the first one to reply

Share this content on your social channels -

Only logged in users can reply.