PaperlayerView exact revision on arXiv

Source-linked AI summary

SiamAPN++: Siamese Attentional Aggregation Network for Real-Time UAV Tracking

Ziang Cao, Changhong Fu, Junjie Ye, Bowen Li, Yiming Li

arXiv:2106.08816v2cs.CV

TL;DR

UAV tracking requires both robust accuracy under fast motion, low resolution, and occlusion and efficiency on embedded platforms. SiamAPN++ combines self- and cross-attentional aggregation with dual-feature anchor proposals, and evaluations report superior benchmark performance plus real-time embedded tracking.

  • Problem

    Existing UAV trackers struggle to combine high performance with the efficiency required by embedded platforms under fast motion, low resolution, and severe occlusion.

  • Method

    SiamAPN++ uses self-AAN and cross-AAN to aggregate feature dependencies and APN-DF to improve anchor robustness across object scales.

  • Results

    SiamAPN++ outperforms its baseline and other SOTA trackers on two UAV benchmarks, while embedded-platform tests demonstrate real-time tracking speed.

  • Takeaways & Limitations

    The proposed tracker provides a practical attentional Siamese approach for real-time UAV tracking with reported accuracy, robustness, and embedded-platform applicability.

Abstract

from arXiv · show

Recently, the Siamese-based method has stood out from multitudinous tracking methods owing to its state-of-the-art (SOTA) performance. Nevertheless, due to various special challenges in UAV tracking, \textit{e.g.}, severe occlusion and fast motion, most existing Siamese-based trackers hardly combine superior performance with high efficiency. To this concern, in this paper, a novel attentional Siamese tracker (SiamAPN++) is proposed for real-time UAV tracking. By virtue of the attention mechanism, we conduct a special attentional aggregation network (AAN) consisting of self-AAN and cross-AAN for raising the representation ability of features eventually. The former AAN aggregates and models the self-semantic interdependencies of the single feature map via spatial and channel dimensions. The latter aims to aggregate the cross-interdependencies of two different semantic features including the location information of anchors. In addition, the anchor proposal network based on dual features is proposed to raise its robustness of tracking objects with various scales. Experiments on two well-known authoritative benchmarks are conducted, where SiamAPN++ outperforms its baseline SiamAPN and other SOTA trackers. Besides, real-world tests onboard a typical embedded platform demonstrate that SiamAPN++ achieves promising tracking results with real-time speed.

I. INTRODUCTION

UAV tracking must balance accuracy and robustness against limited embedded computation and challenges such as fast motion, low resolution, and severe occlusion. SiamAPN++ addresses this gap with attentional aggregation and dual-feature anchors, achieving superior benchmark and real-world results at real-time speed.

  • I. INTRODUCTION: SiamAPN++ targets the conflict between computation-heavy robust trackers and efficient but less accurate approaches in challenging UAV scenes.Embedded platforms impose limited computational resources, while UAV tracking involves fast motion, low resolution, and severe occlusion.
  • I. INTRODUCTION: The attentional aggregation network combines self-AAN and cross-AAN to model within-feature and between-feature dependencies adaptively.Self-AAN uses spatial and channel dimensions, while cross-AAN aggregates interdependencies across different feature maps.
  • I. INTRODUCTION: APN-DF uses dual features to improve anchor anti-interference and robustness under severe scale variation.The design exploits complementary feature levels for anchor generation.
  • I. INTRODUCTION: SiamAPN++ achieves superior performance on two challenging UAV benchmarks, especially under fast motion, low resolution, and severe occlusion.The paper also reports real-world embedded-platform tests with real-time speed.

II. RELATED WORKS

Prior UAV trackers mainly use correlation filters or deep Siamese networks, each facing a practical trade-off between efficiency, robustness, or accuracy. SiamAPN++ builds on anchor-free proposal ideas and attention mechanisms by modeling cross-feature dependencies and strengthening dual-feature anchors.

  • II. RELATED WORKS: Correlation-filter trackers offer low computational complexity but insufficient accuracy and robustness in complex UAV scenes.Their online tracking strategy supports deployment on UAVs but limits practical tracking performance.
  • II. RELATED WORKS: Siamese trackers introduced similarity matching and anchor-based classification and regression, but predefined anchors bring hyper-parameter and sample-imbalance issues.Anchor-free APN addresses these problems but remains insufficient for semantic information variation.
  • II. RELATED WORKS: SiamAPN++ extends attention-based tracking by modeling cross-semantic dependencies while reconstructing APN with dual features for stronger anchor robustness.The design targets representation quality and anti-interference under UAV-specific conditions.

III. METHODOLOGY

SiamAPN++ is built from feature extraction, dual-feature anchor proposal, attentional aggregation, and classification-regression subnetworks. It is designed to address SiamAPN’s sensitivity to semantic variation and limited robustness across object scales.

  • III. METHODOLOGY: The baseline APN generates adaptive anchors from a single feature map, avoiding predefined-anchor hyper-parameters and reducing negative-sample imbalance.SiamAPN++ retains this anchor-free baseline while introducing dual-feature processing.
  • III. METHODOLOGY: SiamAPN’s two stated shortcomings are sensitivity to complex semantic information variation and weak robustness for objects with various scales.These limitations motivate the proposed architectural changes.
  • III. METHODOLOGY: SiamAPN++ consists of a feature extraction network, APN-DF, AAN, and classification-regression network.These four subnetworks define the proposed tracker’s overall architecture.

1) Feature extraction network:

The feature extraction stage uses AlexNet and the final two feature levels, while APN-DF combines complementary semantic and detailed information to generate more robust anchors. Its resulting feature representation also supplies position information for AAN.

  • 1) Feature extraction network:: AlexNet is used as the SiamAPN++ backbone, with feature maps from its last two layers serving the tracking task.The template, search image, and kth-layer output are denoted by Z, X, and ϕk(·), respectively.
  • 1) Feature extraction network:: APN-DF combines high-level semantic features and low-level detailed features to handle objects with different scales.High-level features support classification, while low-level features help distinguish objects accurately.

3) AAN:

AAN enhances feature representations by modeling self-semantic dependencies within individual maps and cross-dependencies between different features, including anchor location information.

  • 3) AAN:: AAN combines self-AAN and cross-AAN to enhance feature-map representation through self-semantic and cross-semantic interdependencies.Self-AAN operates on individual feature maps, while cross-AAN integrates information from different features.
  • 3) AAN:: Self-AAN uses spatial and channel attention to model self-semantic interdependencies within a single feature map.Spatial attention captures spatial relationships, while channel attention uses global average and max pooling before feedforward processing.
  • 3) AAN:: Cross-AAN aggregates dependencies between anchor features and channel-enhanced features, incorporating anchor location information for classification and regression.Its two paths apply channel weighting and element-wise concatenation to combine the feature representations.
  • 3) AAN:: AAN output maps highlight effective information and weaken occlusion interference in real-world tracking tests.The visualization is reported as evidence that AAN supports accurate tracking of occluded objects.

4) Classification and regression network:

The classification and regression network retains the baseline SiamAPN structure while redesigning the regression loss to improve convergence speed and accuracy.

  • 4) Classification and regression network:: The network uses multiclassification branches for anchor IoU classification, ground-truth-region point selection, and center-distance estimation.The three branches address complementary criteria for selecting and evaluating anchors.
  • 4) Classification and regression network:: The redesigned regression loss accelerates convergence by increasing gradients at high-loss positions and decreasing them at low-loss positions.The hyper-parameter α controls positive and negative sample emphasis and is set between 1 and 2.
  • 4) Classification and regression network:: 25 epochs versus 37 epochs: SiamAPN++ reaches convergence faster than the baseline under the redesigned regression loss.The passage attributes the faster convergence to the revised loss formulation.

IV. EVALUATIONS

SiamAPN++ is evaluated on two authoritative UAV tracking benchmarks against 14 additional state-of-the-art trackers under a common Siamese backbone.

  • IV. EVALUATIONS: All evaluated Siamese-based trackers use the same ImageNet-pretrained AlexNet backbone for a more consistent comparison.The common backbone is specified as AlexNet pretrained on ImageNet.

A. Implementation details

The implementation uses standard training and evaluation procedures, and real-world testing measures performance on an embedded NVIDIA Jetson AGX Xavier platform.

  • A. Implementation details: SiamAPN++ is trained on COCO, ImageNet VID, GOT-10K, and Youtube-BB using SGD with minibatches of 220 image pairs.The training setup also fine-tunes the last three backbone layers initially with a learning rate of 5×10^-4.
  • A. Implementation details: Around 35 FPS without TensorRT acceleration: SiamAPN++ runs in real-world tests on an NVIDIA Jetson AGX Xavier embedded platform.The tests validate the tracker’s reported accuracy and robustness alongside real-time speed.
  • A. Implementation details: Figure 4 compares overall tracker performance on UAV20L and UAV123@10fps, while Table I reports average attribute-based evaluation across both benchmarks.The figure caption states that SiamAPN++ achieves superior performance against other state-of-the-art trackers.
  • A. Implementation details: Evaluation uses one-pass precision and success-rate metrics, with success ranked by the AUC of the IoU-based success plot.Precision is based on the percentage of frames whose center-location error is below a threshold.

C. Evaluation on UAV benchmarks

SiamAPN++ outperforms competing trackers on UAV20L and UAV123@10fps, with strong benchmark performance and a tuned w1 value of 1.2.

  • SiamAPN++ is reported to improve over other SOTA trackers on two well-known UAV benchmarks.
  • 2.0% higher precision and 4.0% higher AUC than the second-best tracker on UAV20L establish SiamAPN++ as the benchmark leader.UAV20L evaluates long-term tracking across 20 sequences.
  • The tracker’s best overall UAV20L performance occurs at w1 = 1.2, which is used in all experiments.
  • SiamAPN++ achieves 0.764 precision and 0.580 AUC on UAV123@10fps under severe object variation.The benchmark uses 123 sequences recorded at 10 FPS.

2) Attribute-based performance:

Attribute-based evaluations show that SiamAPN++ is particularly robust to camera motion, fast motion, and occlusion, while component analysis links gains to APN-DF and AAN.

  • Attribute-based performance: SiamAPN++ exceeds the second-best tracker under full occlusion by 5.5% in precision and 16.2% in AUC.It also improves AUC by 5.0% in camera-motion scenarios and surpasses the baseline under fast motion.
  • Attribute-based performance: The evaluation examines camera motion, fast motion, full occlusion, partial occlusion, and scale variation across two UAV benchmarks.
  • Ablation study: APN-DF surpasses the SiamAPN baseline and improves tracking of objects with various scales, while AAN further improves accuracy through feature-interdependency aggregation.
  • Parameter analysis: The AUC and precision of SiamAPN++ are best when w1 = 1.2, the value used for evaluation and real-world tests.

5) Qualitative Evaluation:

Qualitative evaluations cover challenging UAV sequences and embedded-platform tests, where SiamAPN++ is reported to maintain strong tracking despite motion blur and occlusion.

  • Qualitative Evaluation: SiamAPN++ achieves superior qualitative tracking across sequences involving scale variation, aspect-ratio change, fast motion, partial occlusion, camera motion, and out-of-view events.The reported improvement is attributed to combining APN-DF and AAN.
  • Real-world tests: During embedded-platform tests, SiamAPN++ can re-determine an object after errors caused by severe camera-motion blur and achieve long-term tracking.The cited test includes severe camera motion, partial occlusion, and similar objects.
  • Conclusion: The conclusion frames self-AAN, cross-AAN, and dual-feature integration as supporting both real-time feasibility and tracker practicability.
Loading 2106.08816v2…