Autonomous vehicles are self-driving vehicles that can recognize their surroundings, understand their risks, and plan their routes. Currently, the sensors that autonomous vehicles use to recognize their surroundings include cameras, radar, and LiDAR. Different from human eyes and cameras, radar and LiDAR do not provide dense information; they only provide sparse information. Therefore, it is difficult to recognize pedestrians and vehicles using radar and LiDAR. It is only possible to calculate the distance between a vehicle and its object by using these sensors. Recent camera-based obstacle detection technology can detect pedestrians and vehicles under good weather conditions, but cannot detect pedestrians and vehicles properly under hostile weather conditions such as heavy snow, heavy rain, or darkness at night.



The laboratory led by Prof. Jae Wook JEON developed a new Deep Learning technology that could recognize and detect pedestrians and vehicles in hostile conditions. First, new technologies of stereo matching and local patterns in hostile environments have been developed for obtaining the robust distance and shape information of objects by using two cameras and the robust feature information, respectively. Then, the resultant information from the camera image information has been used to train the Deep Learning system.

Unlike existing Deep Learning systems which use only a single camera, a new Deep Learning system has been trained by using two cameras installed in a real vehicle. These two cameras have been used to collect training image information, including hostile conditions.

Our lab also developed a Deep Learning technology that can accurately recognize pedestrians, vehicles, traffic lights, and traffic signs on actual roads in real time. With this Deep Learning technology, our lab won 1st prize in the embedded system sector and 2nd prize in the PC sector at the 13th Hyundai Motor Group Future Vehicle Technology Competition: Autonomous Vehicle Competition Image Recognition Area in May 2017.

In future autonomous vehicles, it will be necessary to perform the recognition of lanes and their branch merge locations, recognition of traffic lights and traffic signs, autonomous parking, and context understanding of vehicle environments by using several cameras installed in a vehicle. Our lab will continue intelligent image processing research in order to achieve this goal.