Cooperative Vehicle Perception and Localization Using Infrastructure-based Sensor Nodes

Abstract

Reliable and accurate Perception and Localization (PL) are necessary for safe intelligent transportation systems. The current vehicle-based PL techniques in autonomous vehicles are vulnerable to occlusion and cluttering, especially in busy urban driving causing safety concerns. In order to avoid such safety issues, researchers study infrastructure-based PL techniques to augment vehicle sensory systems. Infrastructure-based PL methods rely on sensor nodes that each could include camera(s), Lidar(s), radar(s), and computation and communication units for processing and transmitting the data. Vehicle to Infrastructure (V2I) communication is used to access the sensor node processed data to be fused with the onboard sensor data. In infrastructure-based PL, signal-based techniques- in which sensors like Lidar are used- can provide accurate positioning information while vision-based techniques can be used for classification. Therefore, in order to take advantage of both approaches, cameras are cooperatively used with Lidar in the infrastructure sensor node (ISN) in this thesis. ISNs have a wider field of view (FOV) and are less likely to suffer from occlusion. Besides, they can provide more accurate measurements since they are fixed at a known location. As such, the fusion of both onboard and ISN data has the potential to improve the overall PL accuracy and reliability. This thesis presents a framework for cooperative PL in autonomous vehicles (AVs) by fusing ISN data with onboard sensor data. The ISN includes cameras and Lidar sensors, and the proposed camera Lidar fusion method combines the sensor node information with vehicle motion models and kinematic constraints to improve the performance of PL. One of the main goals of this thesis is to develop a wind induced motion compensation module to address the problem of time-varying extrinsic parameters of the ISNs. The proposed module compensates for the effect of the motion of ISN posts due to wind or other external disturbances. To address this issue, an unknown input observer is developed that uses the motion model of the light post as well as the sensor data. The outputs of the ISN, the positions of all objects in the FOV, are then broadcast so that autonomous vehicles can access the information via V2I connectivity to fuse with their onboard sensory data through the proposed cooperative PL framework. In the developed framework, a KCF is implemented as a distributed fusion method to fuse ISN data with onboard data. The introduced cooperative PL incorporates the range-dependent accuracy of the ISN measurements into fusion to improve the overall PL accuracy and reliability in different scenarios. The results show that using ISN data in addition to onboard sensor data improves the performance and reliability of PL in different scenarios, specifically in occlusion cases.