N-fold Bernoulli probability based adaptive fast-tracking algorithm and its application to autonomous aerial refuelling

Recently, deep learning has been widely utilized for object tracking tasks. However, deep learning encounters limits in tasks such as Autonomous Aerial Refueling (AAR), where the target object can vary substantially in size, requiring high-precision real-time performance in embedded systems. This paper presents a novel embedded adaptiveness single-object tracking framework based on an improved YOLOv4 detection approach and an n-fold Bernoulli probability theorem. First, an Asymmetric Convolutional Network (ACNet) and dense blocks are combined with the YOLOv4 architecture to detect small objects with high precision when similar objects are in the background. The prior object information, such as its location in the previous frame and its speed, is utilized to adaptively track objects of various sizes. Moreover, based on the n-fold Bernoulli probability theorem, we develop a filter that uses statistical laws to reduce the false positive rate of object tracking. To evaluate the efficiency of our algorithm, a new AAR dataset is collected, and extensive AAR detection and tracking experiments are performed. The results demonstrate that our improved detection algorithm is better than the original YOLOv4 algorithm on small and similar object detection tasks; the object tracking algorithm is better than state-of-the-art object tracking algorithms on refueling drogue tracking tasks.     

GroupNet: Learning to group corner for object detection in remote sensing imagery

Due to the attractive potential in avoiding the elaborate definition of anchor attributes,anchor-free-based deep learning approaches are promising for object detection in remote sensing imagery. Corner Net is one of the most representative methods in anchor-free-based deep learning approaches. However, it can be observed distinctly from the visual inspection that the Corner Net is limited in grouping keypoints, which significantly impacts the detection performance. To address the above problem, ...     

在轨高效目标检测加速技术

针对深度卷积网络目标检测算法参数量大、计算量大以及受星上计算资源、存储资源及功耗的限制，难以实现在轨部署的问题，提出了一种在轨高效目标检测算法加速框架与实现方法。首先，设计了一种可以同时兼容三种卷积算子的计算引擎，有效提高了资源利用率;其次，从通道和卷积核两个维度将目标检测算法模型展开，实现了加速器的高度并行化和可扩展性;最后，在多种FPGA平台上实现了该加速器并对其性能进行了评估。实验结果表明：所提出的加速器计算性能可以达到1843.2 GFLOPs(每秒千兆次浮点运算)，推理时间为0.22 ms。与同类加速器方案相比，所提出的加速器框架在性能、功耗、能效比及推理时间方面具有很大优势，适合部署在资源受限环境中，具有良好的星上应用前景和价值。     

Target localization method of non-cooperative spacecraft on on-orbit service

With the explosion of the number of meteoroid/orbital debris in terrestrial space in recent years, the detection environment of spacecraft becomes more complex. This phenomenon causes most current detection methods based on machine learning intractable to break through the two difficulties of solving scale transformation problem of the targets in image and accelerating detection rate of high-resolution images. To overcome the two challenges, we propose a novel non-cooperative target detection method using the framework of deep convolutional neural network.Firstly, a specific spacecraft simulation dataset using over one thousand images to train and test our detection model is built. The deep separable convolution structure is applied and combined with the residual network module to improve the network’s backbone. To count the different shapes of the spacecrafts in the dataset, a particular prior-box generation method based on K-means cluster algorithm is designed for each detection head with different scales. Finally, a comprehensive loss function is presented considering category confidence, box parameters, as well as box confidence. The experimental results verify that the proposed method has strong robustness against varying degrees of luminance change, and can suppress the interference caused by Gaussian noise and background complexity. The mean accuracy precision of our proposed method reaches 93.28%, and the global loss value is 13.252. The comparative experiment results show that under the same epoch and batchsize, the speed of our method is compressed by about 20% in comparison of YOLOv3, the detection accuracy is increased by about 12%, and the size of the model is reduced by nearly 50%.     

SRS-Net: Training object detectors from scratch for remote sensing images without pretraining

Most of the current object detection algorithms use pretrained models that are trained on ImageNet and then fine-tuned in the network, which can achieve good performance in terms of general object detectors. However, in the field of remote sensing image object detection, as pretrained models are significantly different from remote sensing data, it is meaningful to explore a train-from-scratch technique for remote sensing images. This paper proposes an object detection framework trained from scratch, SRS-Net, and describes the design of a densely connected backbone network to provide integrated hidden layer supervision for the convolution module. Then, two necessary improvement principles are proposed: studying the role of normalization in the network structure, and improving data augmentation methods for remote sensing images. To evaluate the proposed framework, we performed many ablation experiments on the DIOR, DOTA, and AS datasets. The results show that whether using the improved backbone network, the normalization method or training data enhancement strategy, the performance of the object detection network trained from scratch increased. These principles compensate for the lack of pretrained models. Furthermore, we found that SRS-Net could achieve similar to or slightly better performance than baseline methods, and surpassed most advanced general detectors.     

The adaptive Gaussian mixtures unscented Kalman filter for attitude determination using light curves

The Adaptive Gaussian Mixtures Unscented Kalman Filter (AGMUKF) is introduced to estimate the attitude of a Resident Space Object using light curves. This filter models the state probability density function as a Gaussian Mixture. This enables to capture the non-linearities of the light-curve measurement model. A non-linearity index is used to refine the mixture when necessary, and individual Gaussian kernels are merged back together when their relative distance is below a certain threshold. A conventional attitude Unscented Kalman Filter (UKF) is used to propagate and update each kernel. The AGMUKF efficiently maintains the mixture population as low as possible, while still being able to represent non-symmetric, multimodal, arbitrarily complex distributions. Therefore, it is presented as a promising alternative to Particle-Filter-based implementations, the current state of the art used in sequential attitude estimation from light curves. The non-linearity index has been used to show that the measurement model is the main contributor to the system non-linearity. Results have demonstrated the superiority of the AGMUKF w.r.t. the UKF for attitude determination, and that it can converge for high initial state uncertainty cases, successfully capturing the non-Gaussian probability distribution of the state.     

基于改进SSD算法的航拍目标检测算法研究

研究基于深度学习技术的无人机航拍图像目标检测算法，首先介绍目标检测算法SSD（Single Shot MultiBox Detector），并对其特征提取网络进行改进，采用稠密特征提取网络替换原网络的主干特征提取网络，提高算法的特征提取能力，从而提升了算法的检测精度。针对网络实时性问题，在算法中引入分组卷积，极大地减少了网络参数量，提升了网络推理速度。为解决训练中出现的正负样本不均衡问题，利用焦点损失（Focal Loss）改进了原算法的损失函数，进一步提升了网络的收敛速度和精度。最后，通过仿真验证了改进算法在目标检测精度上的优越性。