Multi-Object Detection, Pose Estimation and Tracking in Panoramic Monocular Imagery for Autonomous Vehicle Perception

Abstract

While active sensing such as radars, laser-based ranging (LiDAR) and ultrasonic sensors are nearly ubiquitous in modern autonomous vehicle prototypes, cameras are more versatile because they are nonetheless essential for tasks such as road marking detection and road sign reading. Active sensing technologies are widely used because active sensors are, by nature, usually more reliable than cameras to detect objects, however they are lower resolution, break in challenging environmental conditions such as rain and heavy reflections, as well as materials such as black paint. Therefore, in this work, we focus primarily on passive sensing technologies. More specifically, we look at monocular imagery and to what extent, it can be used as replacement for more complex sensing systems such as stereo, multi-view cameras and LiDAR.

Whilst the main strength of LiDAR is its ability to measure distances and naturally enable 3D reasoning; in contrast, camera-based object detection is typically restricted to the 2D image space. We propose a convolutional neural network extending object detection to estimate the 3D pose and velocity of objects from a single monocular camera. Our approach is based on a siamese neural network able to process pair of video frames to integrate temporal information.

While the prior work has focused almost exclusively on the processing of forward-facing rectified rectilinear vehicle mounted cameras, there are no studies of panoramic imagery in the context of autonomous driving. We introduce an approach to adapt existing convolutional neural networks to unseen 360° panoramic imagery using domain adaptation via style transfer. We also introduce a new synthetic evaluation dataset and benchmark for 3D object detection and depth estimation in automotive panoramic imagery.

Multi-object tracking-by-detection is often split into two parts: a detector and a tracker. In contrast, we investigate the use of end-to-end recurrent convolutional networks to process automotive video sequences to jointly detect and track objects through time. We present a multitask neural network able to track online the 3D pose of objects in panoramic video sequences. Our work highlights that monocular imagery, in conjunction with the proposed algorithmic approaches, can offer an effective replacement for more expensive active sensors to estimate depth, to estimate and track the 3D pose of objects surrounding the ego-vehicle; thus demonstrating that autonomous driving could be achieved using a limited number of cameras or even a single 360° panoramic camera, akin to a human driver perception.