TU-TESL

• 22

According to foreign media reports, in the next few years, cars will gradually become a popular means of transportation. Until then, however, researchers need to develop tools to ensure that such vehicles are safe and efficient in densely populated environments. Because self driving vehicles need to be able to bypass static and moving obstacles, they should have the ability to detect objects quickly and avoid them. One way to achieve this goal is to develop a model that can predict the future behavior of objects or people on the street to estimate the location of objects or people when vehicles approach. However, it is very challenging to predict the future changes in the urban environment, especially when it is necessary to predict human behavior, such as pedestrian behavior or unexpected behavior. Last year, an autopilot of Uber killed a 49 year old woman named Elaine Herzberg in Arizona. The accident and dozens of other accidents caused a lot of discussion about the safety of autopilot and whether such vehicles should be tested in densely populated environments. Recently, the National Transportation Safety Board (NTSB) issued a new document saying that the automatic driving vehicle that caused fatal accidents last year did not identify Herzberg as pedestrians. The report also showed that the autopilot car in the accident had never been trained by pedestrians for pedestrian crossing. At the time of the accident, Herzberg was crossing the road at random. The NTSB report said that Uber's software was defective and did not find Herzberg, which eventually led to his death. The latest analysis released by NTSB may suspend the autopilot project of the company. After the accident, the project has been suspended for several months. In December 2018, it began testing autopilot again. Such new findings show that more advanced AI and more reliable software are needed before autopilot can be tested on real roads. Interestingly, a few days before NTSB released the above documents, researchers from the advanced technologies group, the University of Toronto and UC Berkeley, the Uber autonomous driving team, published a paper in advance, introducing a new technology that can predict pedestrian behavior, called the discrete residual flow network (drf-net). According to the researchers, the neural network can predict the future behavior of pedestrians and capture the inherent uncertainty when predicting long-distance actions. The senior high school entrance examination is believed to be able to predict the future position of pedestrians through the classification distribution of the representative space, and then use such distribution to plan and optimize the path of the autopilot vehicle, taking into account the location of the pedestrian expected. First of all, the researchers introduced that dtf-net network will grid the road map image, that is, it will be converted into the image composed of discrete pixels, and the behavior of pedestrians will be encoded into the grid image of aerial view, which corresponds to the detailed semantic map. Subsequently, the network extracts features from raster images that are particularly useful for predicting pedestrian behavior. Finally, the researchers trained the model to predict the future behavior of pedestrians. A large data set was used to train and evaluate the neural network. The data set contains real world records collected from several cities in North America, including object annotation and online detection trajectory. Such records include the pedestrian tracks manually marked by researchers in 360 degree and 120 meter field of view by vehicle borne lidar sensors. In the assessment conducted by the researchers, dtf-net technology performed well and was superior to other baseline methods in predicting pedestrian behavior. Therefore, this method may help to improve the performance of the automatic driving vehicle, so that it can predict pedestrian actions and plan paths accordingly. Interestingly, the pedestrian behaviors processed and predicted by dtf-net include "not crossing the road disorderly", "crossing the road disorderly", "crossing the crosswalk" and crosswalk. This seems ironic. Recently, the document issued by NTSB pointed out that when a car crash occurred in Arizona, it was impossible to detect pedestrians crossing the road.