URTrack: End-to-end Multi-object Tracking with Transformer

At present, the two-stage multi-object tracking method has made significant progress in general scenes. However, in unstructured road environment, especially in complex scenes such as open pit mining area (such as target occlusion, similar appearance interference and noise in harsh environment), the two-stage model detection box is easy to be lost or mislocated. Motion prediction based on this in the association phase, such as Kalman filtering, can diverge rapidly, leading to ID switching or lost tracking. To address this deficiency, we analyze the adaptability of the end-to-end model in complex mining scenarios, and propose URTrack tracking framework. Our tracking framework performs object detection and explicit association in an alternating fashion, locking on the evolution of a single object instance by propagating tracking queries across frames, while generating decoupled detection queries to capture all potential objects in the current frame. At the same time, we introduce a learning data association module based on edge-enhanced crossattention to construct a more comprehensive feature dimension. We used the real open-pit mining environment data set collected by ourselves for experimental verification. The data set contains multiple complex scenes, and the amount of data and scene are rich, which is in line with the characteristics of unstructured roads. Experimental results show that URtrack is significantly better than traditional methods in key indicators such as AMOTA and AMOTP, which is improved by 20 percentage points, effectively improving the tracking performance of autonomous driving perception system in complex unstructured environment. This study provides a new solution and technical idea for robust multi-target tracking under harsh conditions.