Detection and Characterisation of Partially Occluded Vulnerable Road Users

Accurate detection and classification of vulnerable road users (pedestrians, cyclists, and micro-mobility users) is a safety critical requirement for the deployment
of autonomous vehicles in heterogeneous trac. Object detection systems have improved significantly in recent years with the proliferation of deep learning-based solutions and the availability of larger and more diverse datasets. Despite this, many
challenges still exist before the detection capabilities required for safe autonomous
driving can be achieved. One of the most complex and persistent challenges is that of
partial occlusion, where a target object is only partially available to the sensor due to
obstruction by another foreground object. The frequency and variety of occlusion in
the automotive environment is large and diverse as pedestrians, e-scooter riders and
cyclists navigate between vehicles, buildings, trac infrastructure and other road
users. Vulnerable road users can be occluded by static or dynamic objects, may
inter-occlude (occlude one another) such as in crowds, and self-occlude - where parts
of a pedestrian or cyclist overlap. This thesis provides in-depth analysis into this
complex object detection challenge and makes significant contributions to the field
of research for partially occluded vulnerable road user detection.
The research identifies a number of knowledge gaps and provides advanced characterisation tools to improve the analysis of state of the art pedestrian and e-scooter
rider detection models. A thorough literature review of occlusion handling techniques
for vehicle detection, vulnerable road user detection and object detection in the automotive environment is presented. A novel, objective metric and methodology for
pedestrian occlusion level classification for ground truth annotation is described that
more accurately reflects the pixel wise occlusion level than the current state of the
art. Two novel, objective test datasets are presented for benchmarking pedestrian and e-scooter rider detection performance for the complete range of occlusion levels
from 0-99%. Finally, a novel occlusion-aware method of e-scooter rider detection is
described that provides a 15.93% improvement over the current state of the art.