Google cars will recognize pedestrians with high accuracy
Before the advent of unmanned vehicles on sale, there is not much time left. Google brings to mind the technology of machine vision and one of the most important elements - pedestrian recognition in real time.
Google Research researchers Alex Krizhevsky, Anelia Angelova and colleagues presented a new method for detecting pedestrians using neural networks . The method showed an encouraging result: a pedestrian can be detected with high accuracy in 73.8% of cases by the Caltech Pedestrian test, which contains a database of 50,000 marked pedestrians in urban and rural conditions. This result is comparable with the best alternative designs.
Experts say that one of the promising areas in recent years has becomethe use of GPU accelerators in neural networks to execute almost real-time complex algorithms for pattern recognition, NLP and fluent analysis of the video stream.
At the recent Nvidia GPU Technology Conference, several such solutions were introduced . Technological progress in this area is best seen in the ImageNet Large Scale Visual Recognition Challenge test , which has been held since 2010. Since then, the level of errors in the classification of images by machine vision algorithms has dramatically decreased.
Pedestrian recognition is one of the practical tasks where new technology is used. In the future, neural networks with GPU acceleration are likely to find application in surveillance cameras, vehicle traffic control systems, etc.
Returning to the development of Google, recognition of pedestrians is a very difficult task, because it needs to be solved in constantly changing environmental conditions in almost real time. Objects around the car are in motion. Existing technologies relatively successfully solve this problem. For example, one of the methods shows an accuracy of 58% in the Caltech Pedestrian test. Another method called VeryFast provides video shooting at 100 frames / s (for comparison, Google shoots at 15 frames / s), but the accuracy is lower there. There are methods with higher accuracy, but they work much slower, reducing speed to 195 times.
Google Research has set a goal to improve recognition accuracy without sacrificing speed. At 15 frames / s, they showed a dramatic increase in accuracy to 73.8% .
The phrase “26.2% average miss rate” from a scientific paper should not be misleading: we are not talking about the fact that the car missed a pedestrian and did not earn points like in Carmageddon. On the contrary, the phrase means that this algorithm “missed” and did not recognize the person to brake in front of him.
At Google Research, the machine vision system was launched on the old-generation Nvidia K20 Tesla GPU graphics accelerator. Now released new versions of K40 and K80, which are used in some supercomputers from the Top500 rating.
The Google Research team intends to improve the result by increasing the depth of the cascade of neural networks and optimizing the ratio of performance and accuracy.
One way or another, but with the advent of robotic cars on the streets, it is better to dress brighter rather than in camouflage so that the machine vision of the car does not confuse you with the background. On the other hand, such advice can be given even now, when cars are often driven by half-blind, half-drunk and distracted drivers.