Lei feng's network has reported that Google's self-driving cars have been able to go out, but many technologies in the car but never released to the public. Recently, in Seattle organized by the IEEE International Conference on Robotics and automation (ICRA), participants saw a new safety device, the device is very rarely seen, is currently being developed by technology giant Google.
Anelia Angelova was a senior researcher at Google, engaged in research and development work in computer vision and machine learning. She displayed a new pedestrian detection system, which can handle video images alone. For any unmanned vehicles, can identify, track and avoid the pedestrian is a very important feature. Google vehicles filled with laser radar and radio detectors and cameras to ensure that they are able to identify hundreds of metres wide pedestrian.
But a set of sensors are expensive, light on top of a rotating laser radar device would cost nearly $ 10000 (if is a multifunctional device then spend more). Network of Lei feng has also broken down before cost of self-driving cars. If automatic vehicles only use cheap cameras to locate pedestrians, which will significantly reduce costs, and will hopefully soon usher in an era of freedom machines to control the car. But the cameras have their own problems. Angelova said, "compared to radar devices, Visual information to give you a wider field of vision, but the speed is relatively slow. "
At least it used to be like this. Best video analysis system using a deep neural networks, machine learning algorithms can be trained, so as to be able to image information (and other types of data) are very accurate classification. Deep neural network rely on multiple layers, which is located between the input and output layers. For image recognition, input will study pixels of the image feature combinations for the next layer to learn about these features, and through the Middle, gradually formed a more sophisticated relationship. Estimation system for the output layer is responsible for looking at something.
Modern deep Web can be such as face recognition and task of one-upmanship, accuracy rate exceeds 99.5%. Angelova, explains, the traditional deep Web is slow for pedestrian detection, each street divides the image into 100,000 or more small pieces, and then in turn to each of the fragments for analysis. This may take several seconds or even minutes per frame, so that they cannot be used for street patrol. Cars cannot use such networks to timely detection of pedestrians may be found in it the pedestrians, they were already past.
Angelova new high speed detection devices are divided into three separate stages.
Phase I is a deep network, compared to before the thousands of fragments, it can split the image into dozens of pieces, can be used in multiple locations at the same time carried out tests in order to recognize pedestrians.
The second stage is another network, it can to improve the recognition results.
Phase III is a traditional deep Web, it will eventually recognize the results, which are found pedestrians, for delivery.
Due to this slow and accurate network analysis on only a fraction of the potential image, so the whole process will go much faster, about 60 to 100 times faster than a network before. Angelova said, the graphics processor to run and are very similar to Google's driverless car will feedback within about a day of street image. Then, it can be about 0.25 seconds to correctly identify pedestrians. Researchers use a known pedestrian image database, instead of using the Google video of the car, so they can compare the results with the previous network. Hello Kitty Note 4 Case
Angelova admitted: "in fact, are not to be used for actual 0.07 seconds. "In order to take action, driverless cars need at the moment to confirm whether pedestrians. "But this means that if other sensors fail, the new system can make a timely supplement treatment. " Hello Kitty Note 4 Case
With the advent of more powerful processors and neural network size increases, Angelova effect of this feature is expected to significantly improve. She said: "from a broad view of the network, you will be able to feel even more rapid development. "Until we can have when no one is driving a car, its unique rotating laser radar may have disappeared completely.
via ieee
519 votes
GoPro Hero4 Silver
As a camera not only do the camera, holding a bunch of extreme sports videos in the Youtube, further down the hearts of grass when GoPro spelling is not the hardware of the camera itself. Sports culture as the core, and relatively good hardware and accessories, GoPro sports camera is you want to do, "Red Bull", once the GoPro truly "moving camera" yardstick of all latecomers, no matter how cheap the price. Performance highlights, will have to face many problems faced by such domestic beverage Red Bull, users to buy or not to buy you, depends entirely on how much you like "Red Bull".
View details of the voting >>
No comments:
Post a Comment