Autonomous driving has caused extensively attention of academia and industry. Vision-based dangerous object detection is a crucial technology of autonomous driving which detects object and assesses its danger with distance to warn drivers. Previous vision-based dangerous object detections apply two independent models to deal with object detection and distance prediction, respectively. In this paper, we show that object detection and distance prediction have visual relationship, and they can be improved by exploiting the relationship. We jointly optimize object detection and distance prediction with a novel multi-task learning (MTL) model for using the relationship. In contrast to traditional MTL which uses linear multi-task combination strategy, we propose a Cartesian product-based multi-target combination strategy for MTL to consider the dependent among tasks. The proposed novel MTL method outperforms than the traditional MTL and single task methods by a series of experiments. 

 

Multi-task Learning with Cartesian Product-Based Multi-objective Combination for Dangerous Object Detection. Available from: https://www.researchgate.net/publication/318136674_Multi-task_Learning_with_Cartesian_Product-Based_Multi-objective_Combination_for_Dangerous_Object_Detection [accessed Dec 31 2017].