Visible and infrared (RGB-IR) fusion object detection plays an important role in security, disaster relief, etc. In recent years, deep-learning-based RGB-IR fusion detection methods have been developing rapidly, but still struggle to deal with the complex and changing scenarios captured by drones, mainly due to two reasons: (A) RGB-IR fusion detectors are susceptible to inferior inputs that degrade performance and stability. (B) RGB-IR fusion detectors are susceptible to redundant features that reduce accuracy and efficiency. In this paper, an innovative RGB-IR fusion detection framework based on global–local feature optimization, named GLFDet, is proposed to improve the detection performance and efficiency of drone-captured objects. The key components of GLFDet include a Global Feature Optimization (GFO) module, a Local Feature Optimization (LFO) module and a Channel Separation Fusion (CSF) module. Specifically, GFO calculates the information content of the input image from the frequency domain and optimizes the features holistically. Then, LFO dynamically selects high-value features and filters out low-value features before fusion, which significantly improves the efficiency of fusion. Finally, CSF fuses the RGB and IR features across the corresponding channels, which avoids the rearrangement of the channel relationships and enhances the model stability. Extensive experimental results show that the proposed method achieves the best performance on three popular RGB-IR datasets DroneVehicle, VEDAI, and LLVIP. In addition, GLFDet is more lightweight than other comparable models, making it more appealing to edge devices such as drones. The code is available at https://github.com/laochen330/GLFDet.

Drone-based small object detection is of great significance in practical applications such as military actions, disaster rescue, transportation, etc. However, the severe scale differences in objects captured by drones and lack of detail information for small-scale objects make drone-based small object detection a formidable challenge. To address these issues, we first develop a mathematical model to explore how changing receptive fields impacts the polynomial fitting results. Subsequently, based on the obtained conclusions, we propose a simple but effective Hybrid Receptive Field Network (HRFNet), whose modules include Hybrid Feature Augmentation (HFA), Hybrid Feature Pyramid (HFP) and Dual Scale Head (DSH). Specifically, HFA employs parallel dilated convolution kernels of different sizes to extend shallow features with different receptive fields, committed to improving the multi-scale adaptability of the network; HFP enhances the perception of small objects by capturing contextual information across layers, while DSH reconstructs the original prediction head utilizing a set of high-resolution features and ultrahigh-resolution features. In addition, in order to train HRFNet, the corresponding dual-scale loss function is designed. Finally, comprehensive evaluation results on public benchmarks such as VisDrone-DET and TinyPerson demonstrate the robustness of the proposed method. Most impressively, the proposed HRFNet achieves a mAP of 51.0 on VisDrone-DET with 29.3 M parameters, which outperforms the extant state-of-the-art detectors. HRFNet also performs excellently in complex scenarios captured by drones, achieving the best performance on the CS-Drone dataset we built.

Multi-target tracking is facing the difficulties of modeling uncertain motion and observation noise. Traditional tracking algorithms are limited by specific models and priors that may mismatch a real-world scenario. In this paper, considering the model-free purpose, we present an online Multi-Target Intelligent Tracking (MTIT) algorithm based on a Deep Long-Short Term Memory (DLSTM) network for complex tracking requirements, named the MTIT-DLSTM algorithm. Firstly, to distinguish trajectories and concatenate the tracking task in a time sequence, we define a target tuple set that is the labeled Random Finite Set (RFS). Then, prediction and update blocks based on the DLSTM network are constructed to predict and estimate the state of targets, respectively. Further, the prediction block can learn the movement trend from the historical state sequence, while the update block can capture the noise characteristic from the historical measurement sequence. Finally, a data association scheme based on Hungarian algorithm and the heuristic track management strategy are employed to assign measurements to targets and adapt births and deaths. Experimental results manifest that, compared with the existing tracking algorithms, our proposed MTIT-DLSTM algorithm can improve effectively the accuracy and robustness in estimating the state of targets appearing at random positions, and be applied to linear and nonlinear multi-target tracking scenarios.