Table 2 Comparative table of key features of major tracking algorithms.
From: Advanced algorithms for UAV tracking of targets exhibiting start-stop and irregular motion
Tracking Algorithm/Approach | Core Principle/Backbone | Key Features | Strengths | Limitations | Typical Application/Notes | References |
|---|---|---|---|---|---|---|
Adaptive Spatial Regularization Correlation Filters (DTSRT) | Correlation filter with deep features and adaptive spatial regularization | Integrates deep features into correlation filter- Uses HOG for scale, deep features for location- Saliency-based spatial constraint for template | No GPU required- Improved accuracy over classic CF, Real-time tracking on CPU | Still limited by hand-crafted features for scale, less robust than deep trackers | UAV tracking on resource-limited platforms | |
Siamese Network-Based Trackers (e.g., SiamHSFT, SiamFC, SiamRPN++) | Siamese CNN (AlexNet, ResNet) | Template and search image branches, Hierarchical feature fusion, Attention modules (CBAM, triplet attention), Transformer encoder for context | Robust to appearance changes- Real-time on GPU, High tracking speed (e.g., 126.7 fps)- Good for small, fast-moving targets | Template drift is possible, Struggles with severe occlusion or extreme appearance changes | UAV tracking in challenging, real-time scenarios | |
Transformer-Based Trackers (e.g., OSTrack, MCJT) | Unified transformer backbone | One-stream feature extraction and correlation, Candidate elimination for background suppression- Motion constraint and spatial remapping modules- Adaptive template update | Superior context modelling, Robust to occlusion, small targets, dynamic blur- Fast inference, suitable for real-time anti-UAV | May fail if the target leaves the search region. Needs robust template update strategy | Infrared anti-UAV tracking, real-time tracking | |
Hybrid Detection-Tracking Systems (e.g., ADTC) | Detector (YOLOv5) + tracker (Kalman filter, SVM) | Detection-verification, tracking chain, Adaptive candidate selection (scene complexity, SVM)- Kalman filter for motion prediction, Data augmentation for tiny object detection | Handles multiple UAVs- Robust to occlusion, clutter, and small targets, Efficient candidate filtering, scalable to multi-UAV | Complexity in integration, Parameter tuning required- Needs both detection and tracking modules | Multi-UAV, anti-UAV, complex real-world surveillance | |
Particle Flow Filter-Based SLAM | Bayesian filtering, particle flow filter | Nonlinear, non-Gaussian state estimation eliminates particle degeneracy and Is Accurate in high-dimensional state spaces. | Superior accuracy and convergence- Robust to sensor noise- Handles complex navigation scenarios | High computational cost- Real-time issues in high dimensions | UAV navigation, SLAM, state estimation | |
ANFIS-Based Sliding Mode Controller (ANFIS-SMC) | Adaptive neuro-fuzzy inference system + sliding mode control | Combines SMC robustness with ANFIS adaptability, reduces chattering, handles nonlinearities and uncertainties- Real-time adaptive parameter tuning | Improved trajectory tracking under uncertainties- Robust to mass/inertia changes and disturbances, Reduced controller design complexity | Requires training and tuning, computationally more intensive than classic SMC | Fixed-wing UAV trajectory tracking, robust control | |
Deep Reinforcement Learning-Based End-to-End Tracking | Deep neural network with reinforcement learning (SAC) | End-to-end learning from images to control commands, Actor-critic architecture, Reward shaping for distance, direction, and success | Learns complex behaviours, handles rapidly changing targets- Integrates perception and control | Requires extensive training, Sensitive to reward design, may need simulation-to-real transfer | UAV dynamic target tracking, autonomous control |
