3D Multi-Object Tracking (MOT) has achieved tremendous achievement thanks to the rapid development of 3D object detection and 2D MOT. Recent advanced works generally employ a series of object attributes, e.g., position, size, velocity, and appearance, to provide the clues for the association in 3D MOT. However, these cues may not be reliable due to some visual noise, such as occlusion and blur, leading to tracking performance bottlenecks. To reveal the dilemma, we conduct extensive empirical analysis to expose the key bottleneck of each clue and how they correlate with each other. The analysis results motivate us to efficiently absorb the merits among all cues and adaptively produce an optimal tracking manner. Specifically, we present Location and Velocity Quality Learning, which efficiently guides the network to estimate the quality of predicted object attributes. Based on these quality estimations, we propose a quality-aware object association (QOA) strategy to leverage the quality score as an important reference factor for achieving robust association. Despite its simplicity, extensive experiments indicate that the proposed strategy significantly boosts tracking performance by 2.2% AMOTA and our method outperforms all existing state-of-the-art works on nuScenes by a large margin. Moreover, QTrack achieves 51.1%, 54.8% and 56.6% AMOTA tracking performance on the nuScenes test sets with BEVDepth, VideoBEV, and StreamPETR models respectively, which significantly reduces the performance gap between the pure camera and LiDAR-based trackers.

Autonomous driving requires the model to perceive the environment and (re)act within a low latency for safety. While past works ignore the inevitable changes in the environment after processing, streaming perception is proposed to jointly evaluate the latency and accuracy into a single metric for video online perception. In this paper, instead of searching trade-offs between accuracy and speed like previous works, we point out that endowing real-time models with the ability to predict the future is the key to dealing with this problem. We build a simple and effective frame-work for streaming perception. It equips a novel Dual-Flow Perception module (DFP), which includes dynamic and static flows to capture the moving trend and basic detection feature for streaming prediction. Further, we introduce a Trend-Aware Loss (TAL) combined with a trend factor to generate adaptive weights for objects with different moving speeds. Our simple method achieves competitive performance on Argoverse-HD dataset and improves the AP by 4.9% compared to the strong baseline, validating its effectiveness. Our code will be made available at https://github.com/yancie-yjr/StreamYOLO.

Abstract Most anchor‐based detectors use intersection over union (IoU) to assign targets to anchors during training. However, IoU did not pay enough attention to the proximity of the anchor's centre to the centre of the truth box, resulting in two issues: (1) the most slender objects were given just one anchor, resulting in insufficient supervision information for slender objects during training; (2) IoU cannot accurately represent the degree of alignment between the feature's receptive field at the anchor's centre and the object. As a result, some features with good alignment degrees are missing, while others with poor alignment degrees are used, reducing the model's localisation accuracy. To address these issues, we first created a Gaussian Guided IoU (GGIoU), which prioritises the proximity of the anchor's centre to the truth box's centre. We then proposed GGIoU‐balanced learning methods, including GGIoU‐guided assignment strategy and GGIoU‐balanced localisation loss. This method can assign multiple anchors to each slender object, favouring features that are well‐aligned with the objects during the training process. A large number of experiments show that GGIoU‐balanced learning can solve the aforementioned problems and significantly improve the detection model's performance.

The teaching–learning-based optimization (TLBO) algorithm has been one of most popular bio-inspired meta-heuristic algorithms due to the competitive converging speed and high accuracy. A batch of TLBO variants has been proposed to enhance the exploitation ability and accelerate the exploration process. However, they still suffer from premature convergence in solving complex non-linear problems. In the study, a novel TLBO variant named dynamic-opposite learning TLBO (DOLTLBO) is proposed, which employs a new dynamic-opposite learning (DOL) strategy to overcome premature convergence. The search space of DOL has the characteristics of asymmetry and dynamically adjusting along with a random opposite number. The asymmetric search space significantly increase the probability for the population in obtaining the global optimum, which holistically improves the exploitation capability of DOLTLBO. Meanwhile, the dynamically changing characteristic enriches the diversity of the search space, thus enhancing the exploration ability. To validate the proposed DOL operator and DOLTLBO algorithm, shifted and rotated benchmark functions from CEC 2014, multiextremal functions and constrained engineering problems have been experimented upon. Comprehensive numerical results with the comparisons with the state-of-the-art counterparts show that DOLTLBO has significant advantages of converging to the global optimum on most benchmarks and engineering problems, which also validates the superiority of the novel DOL operator.