Deep Learning For Better Pedestrian Detection

We know of automakers’ obsession with driverless cars for future motoring. We also know of Google’s involvement in autonomous vehicles and of Apple’s ‘rumored’ Project Titan. What we seldom hear of are the people working in the background, developing and advancing technologies that could end up in our future vehicles in one form or another.

One such group are the engineers at Jacobs School of Engineering at University of California San Diego who are really good at mathematics. Current crash-avoidance systems on cars and experimental autonomous vehicles have a complex array of sensors, radar, lidar and cameras that keep the cars on the road and avoid crashing into objects and pedestrians. These systems take up space, add to cost and have operational limitations.

The current systems require a combination of sensors because ‘cascade detection’ (which identifies and discards things like the sky, buildings, trees, etc) is sometimes too simple in certain driving situations while the more detailed ‘deep neural network’ system which is suited for complex pattern recognition works too slowly for the real world. As such, UC San Diego researchers are developing a detection system that uses visual cues only, much like the human eye, by combining these two systems.

Sounds easy but it involves developing a new algorithm to make pedestrian detection easier, faster and more accurate. Nuno Vasconcelos, electrical engineering professor at UC San Diego said “We’re aiming to build computer vision systems that will help computers better understand the world around them. 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”.

The new algorithm uses cascade detection in the early stages of detection to weed out the background and non-person objects and then deep learning models in the later stages. The algorithm can now 2 – 4 image frames per second which is near real-time but still not quick enough. It’s also said to have better accuracy with half the error rate of similar systems.

The algorithm is currently restricted to binary detection tasks such as pedestrian detection but the team hopes to extend this system to detect multiple objects simultaneously. “One approach to this problem is to train, for example, five different detectors to recognize five different objects. But we want to train just one detector to do this. Developing that algorithm is the next challenge,” added Vasconcelos.

This technology isn’t just bound to driverless cars but can potentially be applied to manufacturing and military applications. Check out the video of the detection system.