Obstacle detector for level crossing using infrared cameras


Author: Ryuta Nakasone
Co-authors: Hiroki Mukojima, Nozomi Nagamine

Day: Introduction Day
Session: Level Crossings

Level crossings, being the interface between railway and road traffic, compose a potential risk of accidents to railway operation. In Japan, there are around 33,000 level crossings and over 90% are equipped with automatic barriers. Despite these efforts made by the government and railway operators, every year more than 200 people are injured due to accidents inside level crossings. In order to detect pedestrians trapped inside level crossings, we propose a new Obstacle Detector (OD) using infrared camera and image processing algorithm mounted on fail-safe processing unit.ODs have been introduced in Japan since the 1960s. A conventional approach is to use laser beams that detect an obstacle by breaking the loop made by the beam. This is a simple and robust way to detect large obstacles inside the level crossing but has the limit of sensing areas, only a single line made by the laser. Nowadays new ODs using LiDAR or RADAR as a source of detection have been introduced. Although these new sensors are capable of detecting obstacles spatially, its performance to detect obstacles close to the road surface is unstable. This is a problem because, for example, detection of people who have fallen down inside the level crossing becomes difficult. In order to solve such problem and improve detection performance, we have developed new OD using infrared camera. Under the method proposed infrared camera is used as the sensor to acquire constant image of the crossing and to detect targets irrespective of their height. By applying the infrared camera, the device has features not found in the conventional devices, such as excellent resistance against weather and sunlight conditions, and does not require illumination.Obstacle detection is performed by image processing technique, using a combination of background subtraction and machine learning method. In general, background subtraction methods have a weakness of detect static objects, in this case a person stuck inside the crossing for a period of time. To overcome this problem, we apply Convolutional Neural Network (CNN) which allows us to reduce the influence of environmental noise and still to be able to detect static targets.We have developed a prototype system in which the detection algorithms is implemented in a pair of Field-Programmable Gate Array (FPGA) boards with fail-safe architecture consists of bus synchronized dual CPUs. Performance verification tests have been conducted at several locations, and as a result, have confirmed that the device has an ability to detect people with minimum erroneous detection regardless of weather and seasonal factors. Furthermore, we will conduct field tests from December 2018, using the prototype to verify the performance under various conditions and situations.Our paper describes the devices' architecture and image processing algorithms for obstacle detection in the level crossing and introduces the results of experiments. We are sure that our proposing method contributes to improvement of conventional devices and enhance safety of the level crossing.