Vehicle Localization with Surveillance Cameras

Vehicle localization is an important problem in autonomous driving research. Main stream methods use sensors on the vehicle with a SLAM algorithm, yet it has disadvantage on safety, cost and global intelligence comparing to sensing from infrastructure, especially surveillance cameras which are already pervasive in public urban area. To better address these challenges, this paper, for the first time, uses multiple surveillance cameras for vehicle localization. In this paper, a flexible two-stage framework for vehicle localization with surveillance cameras is introduced. Flexible deploying is its advantage. With first stage deploying on the camera local computing and second stage on the cloud, it can runs on limited bandwidth and computing conditions. Two potential directions of solutions under the framework are proposed. One is using instance mask as intermediate information, another is using key points of the vehicle.