Probabilistic Approach for Road-Users Detection

Object detection in autonomous driving applications implies the and tracking of semantic objects that are commonly native to urban environments, as pedestrians vehicles. One major challenges state-of-the-art deep-learning based object false positives which occur with overconfident scores. This is highly undesirable other critical robotic-perception domains because safety concerns. paper proposes an approach alleviate problem predictions by introducing a novel probabilistic layer deep networks testing. The suggested avoids traditional Sigmoid or Softmax prediction often produces predictions. It demonstrated proposed technique reduces overconfidence without degrading performance on true positives. validated 2D-KITTI objection through YOLOV4 SECOND (Lidar-based detector). enables interpretable requirement re-training network therefore very practical.