Hitachi Automotive Systems Ltd said has employed AI (artificial intelligence) for its stereo camera designed for automatic brakes.
The camera realizes a function to detect pedestrians during nighttime hours by using several hundreds of thousands of pieces of data as teacher data. Rival companies are currently at the stage of developing sensors supporting AI. Hitachi Automotive is supplying the new sensor to Suzuki Motor Corp, taking a lead in the commercialization of AI sensor.
The new sensor is much superior to the products of rivals around the globe in terms of performance, too. The performance of an automatic brake supporting nighttime pedestrians realized by the new sensor is higher than the performance of vehicles equipped with "EyeQ3," Israel-based Mobileye's main image processing chip.
Hitachi Automotive's previous camera recognizes objects "based on rules," meaning that developers manually set conditions, as in the case of other companies' products. However, with the rule-based method, conditions become complicated, and it becomes difficult to support the detection of objects during nighttime hours. This time, by using machine learning, it becomes possible to efficiently find conditions in a large amount of data.
In general, stereo cameras detect objects that have desired shapes and are located in front of the user's vehicle by using the parallax between two images taken by right and left cameras, respectively. And methods such as pattern recognition are used to judge whether a detected object is a pedestrian or not. The new stereo camera of Hitachi Automotive uses machine learning for the process of image recognition.
Several hundreds of thousands of pieces of teacher data are stored in the image processing microcomputer of the new camera. Images taken with the camera are compared with the teacher data to judge whether an object is a pedestrian or not. The previous stereo camera of Hitachi Automotive uses a normal pattern recognition method that uses multiple images to make such a judgment.
With the application of machine learning, it became easier to detect a pedestrian even (1) when only the lower body is lighted by a headlight or (2) when the entire body is visible but brightness differs in each part of the body, compared with conventional pattern recognition methods.
When machine learning is applied to image recognition processing, the amount of data to be processed increases. To address this problem, Hitachi Automotive renovated the microcomputer of the stereo camera and improved performance. The previous stereo camera uses three microcomputers for image processing, image recognition and control of vehicle, respectively.
For the new stereo camera, Hitachi Automotive reduced the number of microcomputers to two by integrating microcomputers for image processing and image recognition. Then, the number of cores used for image recognition was increased from one to two. By increasing the number of cores, it became possible not only to apply machine learning but also to increase the speed of image recognition processing.
Furthermore, Hitachi Automotive increased the dynamic range of CMOS sensor and reduced the F value of lens, increasing camera sensitivity by 100%. With the wider dynamic range, it became possible to capture both bright and dark objects. With the smaller F value, it became easier to detect pedestrians in the dark.
- External Link