Multi-Vehicle Tracking Through Occlusions Using a Long Short-Term Memory Approach

Importing recorded real-world data into simulation
environments is essential for evaluating automated
driving functionalities. However, during data acquisition,
objects may be occluded for a short period of time, resulting
in incomplete trajectories. These data gaps present
significant challenges for accurate scenario reconstruction
and subsequent system validation. As a result of these
occlusions, fragmented trajectories are formed, known as
tracklets. The aim of the contribution is to associate the
tracklets that belong to the same vehicle using an long
short-term memory-based model. The proposed network
classifies all feasible tracklet sets into two classes: class
match or class non-match. To this end, two networks with
different structures are considered: the all-tracklet network
and the two-tracklet network. The all-tracklet network processes
tracklet sets simultaneously, while the two-tracklet
network processes the set of tracklets sequentially. To
guarantee one-to-one assignments of tracklets, a Hungarian
algorithm is applied to the network’s output generating the
final tracklet association. Experimental results show that
the proposed networks are able to associate the tracklets
accurately with a precision, recall, and F1-scores of higher
than 0.9 for both classes.