Deep Generic Dynamic Object Detection Based on Dynamic Grid Maps

This paper describes a method to detect generic dynamic objects for automated driving. First, a LiDAR-based dynamic grid is generated online. Second, a deep learning-based detector is trained on the dynamic grid to infer the presence of dynamic objects of any type, which is a prerequisite for safe automated vehicles in arbitrary, edge-case scenarios. The Rotation-equivariant Detector (ReDet) – originally designed for oriented object detection on aerial images – was chosen due to its high detection performance. Experiments are conducted based on real sensor data and the benefits in comparison to classic dynamic cell clustering strategies are highlighted. The false positive object detection rate is strongly reduced by the proposed approach.

I. Introduction

The perception and representation of the environment is a key ingredient in automated driving systems. A multitude of data fusion methods and ways of modeling the local area around the ego vehicle have been proposed [1], [2]. With regard to dynamic objects, the majority of work focuses on the detection of known object classes such as vehicles, cyclists, or pedestrians. Deep neural networks are trained on established labeled datasets such as KITTI or nuScenes based on camera, LiDAR, and RADAR data to detect such predefined object classes [3], [4]. In reality, however, the spectrum of objects that can be dynamic is not limited to predefined classes, but nearly anything can move. Examples include shopping carts, rolling tires, or all kinds of animals. But also standard classes such as vehicles exist in all kinds of non-standard appearances, see Fig. 1. Detectors trained on predefined object classes are incapable to perceive such generic dynamic objects – let alone to estimate their velocities or accelerations, which can lead to dangerous situations.