Master Thesis presentation by Amanda Allgurén and Albin Jansfelt

Examiner: Lennart Svensson

Abstract:
The field of autonomous driving advances by the expanding utilization of deep
learning methods. Training deep learning models on automotive data sequences
allows for predictions of an object’s location and its future movements. Harnessing
the benefits of deep learning generally requires accompanying annotations, and since
the annotation process poses a significant bottleneck, new methods to mitigate this
challenge are urgently needed. To address this issue, we propose an auto-annotation
pipeline consisting of three modules. First, a 3D object detector is trained on
annotated single-frame data and thereafter applied to each frame in a sequence.
Second, a model-based tracker connects the bounding boxes across frames and
improves low-confidence detections via filtering. Third, we introduce a smoothing
network that further refines the detections by also incorporating future frames. The
smoothing network considers both bounding boxes and point clouds. With our
smoothing network, we show an improvement in center point, size and rotational
error. This progress aligns with the efforts of previous work that have developed
pipelines integrating object detection, multi-object tracking, and bounding box
refinements [1], [2]. However, we distinguish ourselves by working offline with limited
sequence annotations. In particular, our pipeline works with sequences where only
one frame is annotated. Additionally, we contribute by proposing random window
slides as a data augmentation approach. Our work serves as a baseline for object
detection, multi-object tracking and smoothing for the Zenseact Open Dataset.

Welcome!
Amanda, Albin and Lennart