2020 · 43 citations · 9 references
ComputingAutomotive EngineeringComputer ScienceAutonomous VehiclesComputer VisionEmbedded SystemsAutomotive Embedded SystemsDeep LearningLocalization UncertaintyObject DetectionObject RecognitionUncertainty EstimationAccuracyMean Average PrecisionObject Detection In Autonomous DrivingObject LocalizationPerception SystemsPrediction For Autonomous DrivingReal-time Object DetectionVehicle Detection
For self-driving cars that operate based on battery-generated power, detection and control are commonly performed in embedded systems to reduce power consumption. To drive safely without driver intervention, it is essential to operate object detection algorithms with high accuracy and fast detection speed within autonomous driving embedded systems. This paper proposes new methods to predict the localization uncertainty by applying Gaussian modeling to the DNN-based tiny YOLOv3 algorithm and consequently, to drastically improve accuracy at the expense of a slight penalty of detection speed by using it in post-processing. Compared to the baseline algorithm (i.e., tiny YOLOv3), the proposed algorithm, tiny Gaussian YOLOv3, improves the mean average precision (mAP) by 2.62 and 4.6 on the Berkeley deep drive (BDD) and KITTI datasets, respectively. Nevertheless, the proposed algorithm is capable of performing real-time detection at 55.56 frames per second (fps) on the BDD dataset and 69.74 fps on the KITTI dataset, respectively, under the mode 0 of the autonomous driving embedded platform, Jetson AGX Xavier.
9
Are we ready for autonomous driving? The KITTI vision benchmark suite
Andreas Geiger, P Lenz, R. Urtasun · 2012 · 14.4K citations
YOLOv3: An Incremental Improvement
Joseph Redmon, Ali Farhadi · arXiv (Cornell University) · 2018 · 5.9K citations · Full text
BDD100K: A Diverse Driving Video Database with Scalable Annotation Tooling.
Fisher Yu, Wenqi Xian, Yingying Chen et al. · arXiv (Cornell University) · 2018 · 754 citations · Full text