Abstract
Online multi-object tracking (MOT) has broad applications in time-critical video analysis scenarios such as advanced driver-assistance systems (ADASs) and autonomous driving. In this paper, the proposed system aims at tracking multiple vehicles in the front view of an onboard monocular camera. The vehicle detection probes are customized to generate high precision detection, which plays a basic role in the following tracking-by-detection method. A novel Siamese network with a spatial pyramid pooling (SPP) layer is applied to calculate pairwise appearance similarity. The motion model captured from the refined bounding box provides the relative movements and aspects. The online-learned policy treats each tracking period as a Markov decision process (MDP) to maintain long-term, robust tracking. The proposed method is validated in a moving vehicle with an onboard NVIDIA Jetson TX2 and returns real-time speeds. Compared with other methods on KITTI and self-collected datasets, our method achieves significant performance in terms of the “Mostly-tracked”, “Fragmentation”, and “ID switch” variables.
Highlights
Advanced driver-assistance systems (ADASs) and autonomous driving have consistently been a popular research area
Tao et al [31] focused on the learning strategy of matching functions, but they had a large gap in handling specific Multi-object tracking (MOT) problems, e.g., occlusion or model update
In the urban traffic intersection, vehicles show different shapes in our view, The traffic flow became smoother on the highway, in which vehicles kept moving in the same direction with typical highway situations, like cruising, overtaking, following, etc
Summary
Advanced driver-assistance systems (ADASs) and autonomous driving have consistently been a popular research area. Observable motion cues are more complicated since new emerging targets and tracked targets always overlap with each other When it comes to onboard moving camera platforms, these conditions deteriorate, and tracking models need to put more computational overhead on real-time performance. The batch tracking system [12,14,17] utilizes a set of detection results collected by temporal sliding windows of whole frames to generate global trajectories Such offline tracking methods perform well in obtaining an optimal, theoretical global solution in partial time snippets, they are not applicable in handling dramatic model changes in online, long-term tracking.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.