Pedestrian Multi-Object Tracking with Bottleneck Transformer and Enhanced Feature Fusion

Abstract

Due to balanced tracking accuracy and speed, Joint-detection-and-embedding (JDE) tracking paradigm has drawn great attention, which employs a single work to predict detection and appearance features simultaneously. Building on a strong baseline CSTrack, we replace the spatial convolutions in the final block of backbone with a Bottleneck Transformer, which models global relationships across objects and reduces the parameters. Besides, we introduce an enhanced feature fusion block with structural re-parameterization technique to augment multi-feature fusion for alleviating the contradiction between detection and identification embedding subtasks and maintaining the inference-time. The results on MOT16 and MOT17 datasets indicate that our method achieves competitive tracking results.