Most existing deep learning-based trajectory prediction algorithms heavily rely on human expertise, involving iterative manual tuning of their architectures and parameters to tailor prediction models for specific tasks or scenarios. This approach is not only complex to implement and inefficient, but also struggles to balance inference speed with prediction accuracy. To address this challenge, this paper innovatively proposes an improved heterogeneous multi-agent trajectory prediction algorithm utilizing graph neural architecture search. This method automatically conducts an end-to-end graph architecture search to obtain an optimal trajectory prediction model. To enhance model interpretability and its heterogeneous awareness of diverse scenarios, we design a physics- and risk-interaction-based guidance mechanism to steer the architecture search process. Furthermore, we construct a novel neural architecture search loss function, SocialMI-Loss, which comprehensively considers multiple factors such as prediction accuracy, driving region semantic constraints, and model complexity. This function is intended to guide the learning of the trajectory predictor, achieving a harmonious balance between accuracy and computational complexity. A comprehensive series of comparative experiments conducted on three large-scale autonomous driving datasets (nuScenes, Argoverse, and ApolloScape) consistently demonstrates the superior performance of our proposed method. Experimental results indicate that our framework achieves performance comparable to current state-of-the-art methods, while its automatically searched architecture remains remarkably lightweight. Our code is available at:https://github.com/Tu5tra/TrajGNAS.

Effective noise suppression is crucial for the subsequent interpretation tasks of SAR imagery. Traditional SAR image processing techniques often overlook the coherent nature of noise, leading to a loss of vital detail during filtering. With advancements in deep-learning, significant strides have been made in image processing. However, existing deep-learning methods do not fully leverage the imaging mechanisms of SAR, resulting in a lack of specificity and interpretability in the filtering process. To balance noise reduction with detail preservation and to address the “black box” issue in filtering, we propose an interpretable filtering method that employs a correlation-based upward search for density peaks. Initially, we develop an MeanShift-Markov Random Fields filter (MS-MRF) that integrates MeanShift with Markov Random Fields (MRF) in the joint spatial-spectral domain, ensuring both correlation and detail preservation; the derivation of the MS-MRF filter is rigorously grounded in mathematical theory. Subsequently, we integrate MS-MRF with convolutional operations in deep-learning to create a novel convolutional filter, Interpretable MS-MRF Convolution (IMMC), which enhances the model’s interpretability, noise reduction capabilities, and detail retention. Extensive experiments demonstrate that our method outperforms State of the art(SOTA) SAR denoising techniques, achieving an average SSIM of over 85.00% and an average PSNR exceeding 35.00dB across synthetic datasets with varying noise levels, showing significant improvements in noise suppression, detail preservation, and interpretability.

Synthetic aperture radar (SAR) target detectors based on deep learning have difficulty finding a good balance between accuracy and speed. Current pruning methods are usually used for backbone consistent pruning and seldom directly for the whole structure of deep learning target detectors; therefore, for edge-end applications, this article proposes a new full-link general automatic pruning algorithm for SAR target detectors, referred to as SARGap. First, SARGap automatically analyzes the network structure by creating a dependency graph, divides the pair-coupled network structure into the same group, and prunes the same channel for the same group of network structures so that the algorithm can be applied to a variety of complex target detectors. Second, an automatic pruning rate search method (APRS) is designed to search for the optimal pruning rate of each group of network structures in the target detector. Finally, to find a good balance between precision and speed in the automatic search of the pruning rate, a multiobjective optimization loss function (MOOL) is constructed as the APRS objective function. A series of experiments based on SSDD and HRSID, two large-scale SAR target detection datasets, are carried out to prove the superiority of this method. Using Yolov5s as the baseline, SARGap can compress parameters by 84.29%/82.86% and flops by 80.50%/81.93% on two datasets with almost no loss of accuracy. In addition, SARGap can be applied to any deep learning target detector and match hardware computing resources to achieve optimal full-link pruning.

In this work, we propose a voxel-based single-stage fine-grained and efficient point cloud 3D object detection algorithm to address the inadequate granularity in point cloud feature extraction tasks and the imbalance between efficiency and accuracy in single-stage point cloud 3D object detection scenarios. We develop a lightweight multibranch cross-sparse convolution network (LMCCN) that is designed to preserve the feature granularity of the original point cloud while achieving enhanced extraction efficiency. Additionally, we introduce a compact fine-grained self-attention augmented bird’s eye view (BEV) feature extraction module (CFSAM). This module aims to further refine BEV features, enabling the acquisition of both locally and globally enhanced features and thereby augmentingthe perceptual capabilities of the constructed model. Without bells and whistles, the proposed method attains excellent performance on many autonomous driving benchmarks, with detection accuracies of up to 81.67% on KITTI, 72.74% on ONCE, and 84.00% on nuScenes. Moreover, it reaches a peak detection speed of 46.08 FPS, effectively balancing accuracy with speed.

Most of the existing deep learning-based SAR target detection algorithms rely on manual experience to repeatedly adjust structure and parameters to design deep learning models that meet specific scenarios or tasks. The implementation of the above methods is complicated, the design efficiency is low, and most of the designed models are local optimal. In addition, edge-oriented applications have relatively limited computing resources, and it is difficult for existing deep learning compression methods to ensure a balance between accuracy and complexity. In this paper, we innovatively propose a hardware-aware SAR target detection algorithm via multi-objective NAS, referred to as SARNas. Our SARNas method takes object detection accuracy and model computational complexity as joint guidance objectives to automatically search for an optimal SAR object detection model end-to-end. Experimental results show that SARNas can automatically search for a better SAR target detector with balanced detection accuracy and computational complexity.

Most of the existing deep learning-based SAR target detection algorithms rely on manual experience to repeatedly adjust structures and parameters to design models suitable for specific scenarios or tasks. The implementation of the above methods is complicated, the design efficiency is low, and it is difficult to ensure the balance between accuracy and complexity. We innovatively propose a hardware-aware SAR target detection algorithm via multiobjective neural architecture search (NAS), referred to as SARNas. First, we design a flexible and efficient search space, a supernet search strategy and a subnet contribution evaluation strategy. Furthermore, we construct a new NAS loss function, called SARMI-Loss, to guide the learning of a SAR object detector that balances accuracy and computational complexity. Our SAR-Nas method can address the resource limitations of edge devices and automatically search for the optimal SAR target detector in an end-to-end manner for any deep learning-based SAR baseline model. A series of comparative experiments on three SAR image object detection datasets (SSDD, HRSID and MSAR) demonstrate the superiority of our method. The experimental results with YOLOV5 as the benchmark model show that the detection accuracy of the target detection networks automatically found by using the SARNas method on the SSDD, HRSID, and MSAR datasets can reach 98.5%, 92.8%, and 91.8% in mean average precision (mAP) with only 2.31M, 1.99M, 2.21M parameters, respectively. The number of model parameters is reduced by 88.9%, 90.46%, and 68.5%, respectively, and the inference speed is increased by 51.6%, 46.1%, and 13.9% without losing accuracy.

To address the challenges in synthetic aperture radar (SAR) ship target detection, this paper proposes a SAR ship small target orientation detector named FADet based on semantic flow feature alignment and Gaussian label matching. First, to solve the feature misalignment problem caused by feature extraction downsampling and residual connections, we introduce the FAM module into FPN, which automatically aligns deep and shallow fine-grained semantics information through semantic flow alignment. Second, due to the scattering characteristics of SAR imaging, the boundary information of SAR targets is not obvious, we combining attention mechanisms design an adaptive boundary enhancement module to enhance the target boundary information. Finally, to solve the problem that small targets have difficulty matching positive samples under IOU rules, we design a label matching strategy based on Gaussian distribution. This matching strategy can still learn regression information when two boxes do not intersect. Based on the SSDD+ and RSDD-SAR datasets, the effectiveness of each module in FADet is verified by ablation experiments. Additionally, through comparison experiments with the latest orientation detection methods, FADet achieves a good compromise between accuracy and inference speed. The AP50 and AP75 on the SSDD+ and RSDD-SAR is 91.03, 59.94 and 90.78, 59.91 respectively, and the FPS is 19.83.

In recent years, deep learning has been widely used in remote sensing, especially in the field of synthetic aperture radar (SAR) image target detection. However, all of these deep learning models continue increasing the network’s depth and width without maintaining a good balance between accuracy and speed. Therefore, in this article, we propose a hybrid representation learning-enhanced SAR target detection algorithm based on the unique features of SAR images from a lightweight perspective called HRLE-SARDet. First, we design a lightweight and scattering feature extraction backbone that is more suitable for SAR image data. Second, for the multiscale feature discrepancy, we design a new multiscale feature fusion neck. Next, to better extract the scattering information from small targets of SAR images and improve the detection accuracy, we design a lightweight hybrid representation learning enhancement module. Finally, to better fit target detection for SAR image datasets, we redesign a more flexible loss function, which allows for an easy adjustment of the importance of polynomial bases according to the target task and dataset. Extensive experimental results on three SAR image ship target datasets (SSDD, AIR-SARShip-2.0, and HRSID) and a newly released large multiclass target SAR dataset (MSAR-1.0) show that our HRLE-SARDet achieves 98.4%, 79.2%, 92.5%, and 88.4% mean average precision (mAP) with only 1.09 M parameters and 2.5 G floating-point operations (FLOPs) on the SSDD, AIR-SARShip-2.0, HRSID, and MSAR-1.0 datasets, respectively, which is an excellent performance.

According to the surveys of the World Health Organization, distracted driving is one of main causes of road traffic accidents. To improve road traffic safety, real-time detection of drivers’ driving behavior is very important for the development of highly reliable Advanced Driver Assistance System (ADAS). At present, the deep learning architecture based on a Convolutional Neural Network (CNN) has disadvantages such as large number of parameters and weak global feature extraction ability. Therefore, this paper proposes an innovative driver distraction detection model based on the fusion of a transformer and a CNN, referred to as FPT, which is the first exploration in the field of driver distraction detection. First, we introduce the latest Twins transformer as a benchmark. Then, we design residual embedding to replace block embedding, which can further integrate the convolutional neural network with Transformer and improve the feature extraction ability. In addition, the Multilayer Perceptron (MLP) module with a large parameter occupancy rate in the original transformer structure is replaced with a lightweight group convolution module to reduce computational complexity. Finally, a cross-entropy loss function for label smoothing is designed to guide network learning with significantly differentiated features. Comparison results on two large-scale driver distraction detection datasets show that the proposed FPT offers a better compromise between computational cost and performance compared to the state-of-the-art CNN and Transformer architectures.

Unlike optical images, synthetic aperture radar (SAR) images have unique characteristics, such as few samples, strong scattering, sparseness, multiple scales, complex interference and background, and inconspicuous target edge contour information. Current SAR target detection algorithms have difficulty in balancing accuracy and speed, and the performance of these algorithms is relatively limited, thus making it difficult to deploy practical applications. To this end, this article proposes AFSar, an innovative anchor-free SAR target detection algorithm based on multiscale enhancement representation learning. First, we introduce the latest anchor-free architecture YOLOX as the basic framework. Second, to reduce the computational complexity of the model and to improve the ability of multiscale feature extraction, we redesigned the lightweight backbone, namely, MobileNetV2S. Furthermore, we propose an attention enhancement PAN module, called CSEMPAN, which highlights the unique strong scattering characteristics of SAR targets by integrating channel and spatial attention mechanisms. Finally, in view of the multiscale and strong sparse characteristics of SAR targets, we propose a new target detection head, namely, ESPHead. ESPHead extracts the features of targets with different scales by using dilated convolution with different dilated rates, so as to enhance the detection ability of the model for targets with different scales. The results of ablation experiments on the SSDD dataset show that the mAP of our algorithm reaches 0.977, while the Flops is only 9.86 G, achieving state of the art.

At present, few-shot object detection research in the field of optical remote sensing images has been conducted, but few-shot object detection in the field of SAR images have rarely been explored. To this end, this paper proposes a lightweight meta-learning-based SAR image few-shot object detection method, which improves the accuracy and speed of SAR image few-shot object detection from a more balanced perspective. First, we introduce the latest FSODM method in optical remote sensing as a benchmark framework. Second, a lightweight meta-feature extractor named DarknetS is designed to enhance the feature representation of SAR images and improve detection timeliness. Furthermore, we build a new aggregation module called AggregationS, which encodes support features and query features into the same feature subspace via a novel transformer encoder. This module design can better extract the correlation and saliency between different classes in the support set, improve the detection accuracy of the query set, and enhance the detection generalization performance of new classes. Finally, we built several real-world SAR image few-shot object detection datasets to verify the effectiveness of the method. Experimental results show that FSODS can achieve a better object detection performance compared to the baseline model under the condition that only a small amount of labelled data is required for new classes of SAR image objects.

Traffic accidents caused by distracted drivers account for a large proportion of traffic accidents each year, and monitoring the driving state of drivers to avoid traffic accidents caused by distracted driving has become a very important research direction. At present, the field of driver distraction detection mainly adopts supervised learning methods, which have problems such as poor generalization ability, large labeling cost, and weak artificial intelligence. This paper is oriented toward driver distraction fine-grained detection and innovatively proposes a new unsupervised deep learning algorithm, which is referred to as UDL, to achieve a more human-like level of intelligence. First, we build a new unsupervised deep learning algorithm; furthermore, we integrate the multilayer perceptron (MLP) architecture to build a new backbone and projection head to strengthen feature extraction capabilities; and finally, a new loss function based on contrast learning and a stop-gradient strategy is designed to guide the model to learn more robust features. The comparison results on large-scale driver distraction detection datasets show that our UDL method can accurately detect driver distraction without labels and exhibits excellent generalization performance with a linear evaluation accuracy of 97.38%; In addition, after fine-tuning with fewer labels, our UDL method can achieve superior performance close to state-of-the-art supervised learning methods, achieving 99.07% accuracy after fine-tuning using only 50% of the labeled data, which greatly reduces the cost and limitations of manual annotation.

In the future, vehicles will be equipped with increasingly advanced interactive intelligent electronic devices, which will induce drivers to conduct secondary tasks, thereby leading to distractions. Therefore, the detection and early warning of driver distraction are essential for improving driving safety and pose an important challenge in intelligent transportation systems. Previous studies used traditional machine learning and deep learning transfer models, which have the disadvantages of complicated and time-consuming manual feature engineering, strong subjectivity, and weak generalization performance. In this paper, we propose a fine-grained detection method for driver distraction based on neural architecture search. First, we design an automatic construction algorithm for deep convolutional neural networks based on neural architecture search, which automatically searches for the optimal deep convolutional neural network architecture without human involvement. In addition, we fuse driver-related multisource perception information, use an automatically constructed deep convolutional neural network to extract high-dimensional mapping features, and implement fine-grained detection of various types of driver distraction states. The results on a large-scale multimodal driver distraction dataset demonstrate that our method can efficiently search an optimal deep convolutional neural network, which can quickly converge, and can accurately detect the considered types of driver distraction states, the average detection accuracy reaches 99.7796%; moreover, it has satisfactory robustness.

A large number of studies have shown that most vehicle collisions are caused by drivers' abnormal operations. To ensure the safety of all people on the road network as much as possible, it is crucial to be able to predict the drivers' driving safety risks in real time. In this paper, we propose a novel cost-sensitive L1/L2-nonnegativity-constrained deep autoencoder network for driving safety risk prediction. Unfortunately, with existing research methods, the size of the sliding time window is too large, the feature extraction is relatively subjective, and class imbalances occur, which leads to low identification accuracy, long prediction times, and poor applicability. We first propose using a three-layer L1/L2-nonnegativity-constrained autoencoder to adaptively search the optimal size of the sliding window and then construct a deep L1/L2-nonnegativity-constrained autoencoder network to automatically extract the hidden features of the driving behaviors. Finally, we build a new L1/L2-nonnegativityconstrained focal loss classifier to predict the driving behaviors under different safety risk levels. The results from the public 100-Car naturalistic driving study dataset indicate that our method can effectively find the optimal window size, reduce the data volume and reconstruction error, and extract more distinctive features. Furthermore, this method effectively curbs the class imbalance, improves the driving safety risk prediction performance, reduces overfitting, shortens the prediction time, and improves the timeliness.