Relative Position Vector Generation with Computer Vision for Vehicle Platooning Applications

Abstract

Many significant advances have been made in autonomous vehicle technology over the recent decades. This includes platooning of heavy trucks. As such, many institutions have created their own version of the basic platooning platform. This includes the California PATH program [1], Japan’s ”Energy ITS” project [2], and Auburn University’s CACC Platform [3]. One thing these platforms have in common is a strong dependence on GPS based localization solutions. Issues arise when the platoon navigates into challenging environments, including rural areas with foliage which might block receptions, or more populated areas which might present urban canyon effects. Recent research focus has shifted to handling these situations through the use of alternative sensors, including cameras. The perception method proposed in this thesis utilizes the You Only Look Once (YOLO) real-time object detection algorithm in order to bound the lead vehicle using both RGB and IR cameras. Two different YOLO variants were evaluated: YOLOv3 and TinyYOLOv3. Monocular range is determined using both the classical pinhole model and virtual horizon ranging model. A bearing model is introduced which uses the range to determine bearing to the lead vehicle. Various combinations of cameras, YOLO models, and ranging models are then tested on heavy duty truck data collected on roads near Auburn University’s NCAT facility. The results shown in this thesis reveal that there is a slight range accuracy advantage to YOLOv3 on both the pinhole camera model and the Virtual horizon model. TinyYolo was shown to have a faster processing speed which would be ideal in highly dynamic situations. Using the results from the on road truck analysis, a real-time implementation was developed using two consumer sedans. During analysis it was discovered that YOLO had momentary lapses in which it would not detect the lead vehicle, and would therefore not be able to provide range and bearing measurements. To address this, a sub-tracking algorithm was developed. The algorithm was developed around established tracking algorithms, and analysis was performed to determine which tracking algorithm was best suited for dynamic vehicle tracking. Additionally, a slight variation of the method was developed which utilized a stereoscopic camera. The sub-tracking algorithm and stereoscopic vehicle detection algorithm were evaluated in several real-time platooning scenarios, in which the following vehicle operated autonomously using lateral control.