DPTrack:Directional Kernel-Guided Prompt Learning for Robust Nighttime Aerial Tracking

TL;DR

DPTrack tackles nighttime aerial tracking by replacing coarse prompt supervision with fine-grained cues derived from the object’s intrinsic attributes. It introduces a three-module framework: Dual Particle Perception (DPP) to capture global-local topology, a Directional Kernel Adaptive Encoder (DKE) to encode topology into a directional kernel, and a Kernel-Guided Prompt (KGP) module to propagate a channel-aware, spatially gated prompt across the search region. The approach yields a directional kernel that points to the object, enabling precise prompting and robust ground-truth localization under low light. Extensive experiments on UAVDark135, NAT2021, NAT2021-L, NAT2024-1, and DarkTrack2021 demonstrate consistent gains over SOTA methods with real-time performance, validating the practical impact for nighttime aerial tracking in challenging environments.

Abstract

Existing nighttime aerial trackers based on prompt learning rely solely on spatial localization supervision, which fails to provide fine-grained cues that point to target features and inevitably produces vague prompts. This limitation impairs the tracker's ability to accurately focus on the object features and results in trackers still performing poorly. To address this issue, we propose DPTrack, a prompt-based aerial tracker designed for nighttime scenarios by encoding the given object's attribute features into the directional kernel enriched with fine-grained cues to generate precise prompts. Specifically, drawing inspiration from visual bionics, DPTrack first hierarchically captures the object's topological structure, leveraging topological attributes to enrich the feature representation. Subsequently, an encoder condenses these topology-aware features into the directional kernel, which serves as the core guidance signal that explicitly encapsulates the object's fine-grained attribute cues. Finally, a kernel-guided prompt module built on channel-category correspondence attributes propagates the kernel across the features of the search region to pinpoint the positions of target features and convert them into precise prompts, integrating spatial gating for robust nighttime tracking. Extensive evaluations on established benchmarks demonstrate DPTrack's superior performance. Our code will be available at https://github.com/zzq-vipsl/DPTrack.

Figures (10)

• Figure 1: Comparison between existing prompt-based trackers and DPTrack. The left illustrates the design philosophy, and the right shows the pipeline. (a): Existing trackers generating prompts solely rely on loss. (b): DPTrack utilizes fine-grained guidance signals to produce accurate prompts.
• Figure 2: Overview of DPTrack. DPTrack utilizes DPP to establish global-local structural correlations, enhancing feature representation; DKE then transforms structured object information into the DK to guide accurate prompt generation. Finally, the KGP quantifies kernel correlations with search features to generate prompts that enable precise object localization.
• Figure 3: The DPP module operates in two stages: LCP captures global structural information for scale-aligned concatenation with the original features, and MCP extracts local details to establish global-to-local structural associations. The fused features are then convolved to produce the association matrix.
• Figure 4: The impact of non-uniform illumination on the score map: (a) The green ground-truth box indicates the object's true position, (b) presents the corresponding score map (artifact from localized glare) and (c) depicts the true score map after interference suppression by IE.
• Figure 5: Overview of the KGP pipeline. Input features are scale-aligned by convolution, and two separate convolutions are applied to adjust $F^{Hr}_X$ for differences in emphasized dimensions, producing a fused representation with $F_X$ that preserves the original feature information. The $DK$ estimates the confidence of each channel in $F^{Hr}_X$ for object indication, and the refined $Sim$, after L2 normalization, is employed as prompt to guide the tracker toward high-confidence channels. Finally, gating units partition the features along the spatial dimension to filter out noise.