Object Detector Research Articles

Overcoming the Limitations of Learning-Based VQA for Counting Questions with Zero-Shot Learning

Visual question answering (VQA) research has garnered increasing attention in recent years. It is considered a visual Turing test because it requires a computer to respond to textual questions based on an image. Expertise in computer vision, natural language processing, knowledge understanding, and reasoning is required to solve the problem of VQA. Most techniques employed for VQA consist of models that are developed to learn the combination of image and question features along with the expected answer. The techniques chosen for image and question feature extraction and combining the features change with each model. This method of teaching a model of the question–answer pattern is ineffective for queries that involve counting and reasoning. This approach also requires considerable resources and large datasets for the training. The general VQA datasets feature a restricted number of items as responses to counting questions ([Formula: see text]), and the distribution of the answers is not uniform. To investigate these issues in VQA, we created synthetic datasets that could be modified to adjust the number of objects in the image and the amount of occlusion. Specifically, a zero-shot learning VQA system was devised for counting-related questions that provide answers by analyzing the output of an object detector and the query keywords. Using synthetic datasets, our model generated 100% correct results. Testing on the benchmark datasets task directed image understanding challenge (TDIUC) and TallyQA-simple indicated that the proposed model matched the performance of the learning-based baseline models. This methodology can be used efficiently for counting VQA questions confined to certain domains when the number of items to be counted is significant.

Voxel self-attention and center-point for 3D object detector

With the advancement of autonomous driving, the industrial demand for 3D object detection has continuously increased, leading to the development of anchor-based LiDAR object detectors reliant on convolutional neural networks (CNN). However, on the one hand, the poor receptive field of CNN limits the understanding of the scene. On the other hand, anchor-based methods cannot accurately predict the posture of objects in the steering. Therefore, in this paper, we propose the voxel self-attention and center-point (VSAC). Firstly, a voxel self-attention network is designed into VSAC to capture extensive voxel relationship. Secondly, considering the impact of feature weight on prediction results, the pseudo spatiotemporal feature pyramid net (PST-FPN) is proposed. Finally, we employ a center-point detection head to make the prediction direction closer to the real object during steering. The experimental results of VSAC on the widely used KITTI dataset, Waymo Open Dataset, and nuScenes dataset demonstrate its positive performance.

Fusion-based modeling of an intelligent algorithm for enhanced object detection using a Deep Learning Approach on radar and camera data

Object detection, the process of detecting and classifying objects within a given environment, forms the foundational element. Multisensory fusion incorporates data from diverse sensors, like radar and cameras, to refine the reliability and accuracy of detection. Further, Radar and camera data fusion refine this process by integrating the unique strength of both technologies, which leverage the radar's proficiency in adverse weather conditions and the camera's high-resolution imaging. This incorporation enhances the object detection systems, which enables them to effectively operate across the spectrum of scenarios, from autonomous vehicles navigating challenging weather to surveillance systems monitoring critical infrastructure. Deep learning (DL), a branch of machine learning (ML), empowers this system with the capability to learn complex representations and patterns directly from the data, which enables them to generalize and adapt to new situations. By integrating the advanced methodology, we can develop strong perception system capable of interpreting and detecting objects accurately in dynamic and diverse environments, from autonomous vehicles navigating urban landscapes to surveillance systems monitoring complex environments. This study designs an Intelligent Algorithm for Enhanced Object Detection Using Deep Learning Approach on the Radar and Camera Data Fusion (IAEOD-DLRCDF) technique. The presented IAEOD-DLRCDF technique uses multi-angle joint calibration where the spatial sparse alignment of the heterogeneous data of the camera and Radar is realized with image falsification disregarded. Besides, the IAEOD-DLRCDF technique applies YOLOv8 object detector for radar and camera target detection individually which are then integrated with the image plane. Moreover, the detected objects are then classified via the bidirectional long short-term memory (BiLSTM) model. Furthermore, the Adam optimizer is used for the optimum hyperparameter selection of the BiLSTM network which results in a better recognition rate. The performance assessment of the IAEOD-DLRCDF method is tested under benchmark dataset. The empirical analysis stated that the IAEOD-DLRCDF method gains better performance over other models.

Cosine similarity-guided knowledge distillation for robust object detectors

This paper presents a Cosine Similarity-Based Knowledge Distillation (CSKD) for robust, lightweight object detectors. Knowledge Distillation (KD) has been effective in enhancing the performance of compact models in image classification by leveraging deep CNN models. However, the complex and multifaceted nature of object detection, characterized by its modular design and multitasking requirements, poses significant challenges for traditional KD techniques. These challenges are further compounded by the conventional reliance on the Mean Squared Error (MSE) loss function and the limited application of enhanced feature representations to the training phase. Addressing these limitations, the proposed CSKD method combines cosine similarity guidance with MSE loss to facilitate a more effective knowledge transfer from the teacher model to the student model. This is achieved by distilling both intermediate features and prediction outputs, aided by an assistant prediction branch designed to learn directly from the teacher’s predictions. This dual-faceted distillation strategy enables the student model to better mimic the teacher model’s behavior, leading to improved performance. The proposed method demonstrates versatility and robustness across various object detector architectures without the need for additional feature enhancement layers during training. Notably, employing ResNet-50 as the teacher model and ResNet-18 as the student model, we achieve new benchmarks in KD for object detection across several architectures, including Faster-RCNN, RetinaNet, FCOS, and GFL, with respective mAP scores of 36.6, 35.2, 35.9, and 38.9. These results highlights the effectiveness of CSKD in advancing the state-of-the-art in KD for object detection, offering a compelling solution to the challenges previously faced by traditional KD methods in this domain. The code of the proposed CSKD is available at https://github.com/swkdn16/CSKD.

Black-box adversarial patch attacks using differential evolution against aerial imagery object detectors

Adversarial patch attacks are commonly used to attack aerial imagery object detectors. However, most existing methods are designed for white-box settings, which are impractical in real world scenarios. In this work, we propose a novel framework for black-box adversarial patch attacks against aerial imagery object detectors using differential evolution (DE). Specifically, we first introduce a new dimensionality reduction strategy to address the high-dimensionality curse in DE optimization problems. Then, we design three universal fitness functions to help DE to select better individuals within a limited computational budget. Finally, we conduct extensive experiments on the Dataset for Object Detection in Aerial images (DOTA) against You Only Look Once (YOLOv3, YOLOv4), and Faster Region-based Convolutional Network (Faster R-CNN), utilizing attack success rate (ASR) and average precision (AP) to quantitatively evaluate the attack efficiency. Results on YOLOv3 for the plane category indicate that our proposed strategy achieves at least a 20.97% improvement in ASR compared to existing dimensionality reduction schemes. Moreover, experiments conducted on detectors with different architectures reveal inconsistent adversarial robustness across categories. To the best of our knowledge, this is the first work to explore the use of DE in black-box adversarial patch attacks against aerial imagery object detectors. Our code has been released at https://github.com/tang-agui/Bbox-Att-Detector.

PS-YOLO: a small object detector based on efficient convolution and multi-scale feature fusion

PS-YOLO: a small object detector based on efficient convolution and multi-scale feature fusion

Transfer learning with generative models for object detection on limited datasets

Abstract The availability of data is limited in some fields, especially for object detection tasks, where it is necessary to have correctly labeled bounding boxes around each object. A notable example of such data scarcity is found in the domain of marine biology, where it is useful to develop methods to automatically detect submarine species for environmental monitoring. To address this data limitation, the state-of-the-art machine learning strategies employ two main approaches. The first involves pretraining models on existing datasets before generalizing to the specific domain of interest. The second strategy is to create synthetic datasets specifically tailored to the target domain using methods like copy-paste techniques or ad-hoc simulators. The first strategy often faces a significant domain shift, while the second demands custom solutions crafted for the specific task. In response to these challenges, here we propose a transfer learning framework that is valid for a generic scenario. In this framework, generated images help to improve the performances of an object detector in a few-real data regime. This is achieved through a diffusion-based generative model that was pretrained on large generic datasets. With respect to the state-of-the-art, we find that it is not necessary to fine tune the generative model on the specific domain of interest. We believe that this is an important advance because it mitigates the labor-intensive task of manual labeling the images in object detection tasks. We validate our approach focusing on fishes in an underwater environment, and on the more common domain of cars in an urban setting. Our method achieves detection performance comparable to models trained on thousands of images, using only a few hundreds of input data. Our results pave the way for new generative AI-based protocols for machine learning applications in various domains, for instance ranging from geophysics to biology and medicine.

A comprehensive survey of deep learning-based lightweight object detection models for edge devices

This study concentrates on deep learning-based lightweight object detection models on edge devices. Designing such lightweight object recognition models is more difficult than ever due to the growing demand for accurate, quick, and low-latency models for various edge devices. The most recent deep learning-based lightweight object detection methods are comprehensively described in this work. Information on the lightweight backbone architectures used by these object detectors has been listed. The training and inference processes concerning to deep learning applications on edge devices is being discussed. To raise readers’ awareness of this developing domain, a variety of applications for deep learning-based lightweight object detectors and related utilities have been offered. Designing potent, lightweight object detectors based on deep learning has been suggested as a counter to such problems. On well-known datasets such as MS-COCO and PASCAL-VOC, we thoroughly examine the performance of certain conventional deep learning-based lightweight object detectors.

Global remote feature modulation end-to-end detection

Object detector based on fully convolutional network achieves excellent performance. However, existing detection algorithms still face challenges such as low detection accuracy in dense scenes and issues with occlusion of dense targets. To address these two challenges, we propose an Global Remote Feature Modulation End-to-End (GRFME2E) detection algorithm. In the feature extraction phase of our algorithm, we introduces the Concentric Attention Feature Pyramid Network (CAFPN). The CAFPN captures direction-aware and position-sensitive information, as well as global remote dependencies of features in deep layers by combining Coordinate Attention and Multilayer Perceptron. These features are used to modulate the front-end shallow features, enhancing inter-layer feature adjustment to obtain comprehensive and distinctive feature representations.In the detector part, we introduce the Two-Stage Detection Head (TS Head). This head employs the First-One-to-Few (F-O2F) module to detect slightly or unobstructed objects. Additionally, it uses masks to suppress already detected instances, and then feeds them to the Second-One-to-Few (S-O2F) module to identify those that are heavily occluded. The results from both detection stages are merged to produce the final output, ensuring the detection of objects whether they are slightly obscured, unobstructed, or heavily occluded. Experimental results on the pig detection dataset demonstrate that our GRFME2E achieves an accuracy of 98.4%. In addition, more extensive experimental results show that on the CrowdHuman dataset, our GRFME2E achieves 91.8% and outperforms other methods.

Attentive context and semantic enhancement mechanism for printed circuit board defect detection with two-stage and multi-stage object detectors

Printed Circuit Boards (PCBs) are key devices for the modern-day electronic technologies. During the production of these boards, defects may occur. Several methods have been proposed to detect PCB defects. However, detecting significantly smaller and visually unrecognizable defects has been a long-standing challenge. The existing two-stage and multi-stage object detectors that use only one layer of the backbone, such as Resnet’s third layer (C4\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$C_4$$\\end{document}) or fourth layer (C5\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$C_5$$\\end{document}), suffer from low accuracy, and those that use multi-layer feature maps extractors, such as Feature Pyramid Network (FPN), incur higher computational cost. Founded by these challenges, we propose a robust, less computationally intensive, and plug-and-play Attentive Context and Semantic Enhancement Module (ACASEM) for two-stage and multi-stage detectors to enhance PCB defects detection. This module consists of two main parts, namely adaptable feature fusion and attention sub-modules. The proposed model, ACASEM, takes in feature maps from different layers of the backbone and fuses them in a way that enriches the resulting feature maps with more context and semantic information. We test our module with state-of-the-art two-stage object detectors, Faster R-CNN and Double-Head R-CNN, and with multi-stage Cascade R-CNN detector on DeepPCB and Augmented PCB Defect datasets. Empirical results demonstrate improvement in the accuracy of defect detection.

Deep-learning assisted detection and quantification of (oo)cysts of Giardia and Cryptosporidium on smartphone microscopy images

The consumption of microbial-contaminated food and water is responsible for the deaths of millions of people annually. Smartphone-based microscopy systems are portable, low-cost, and more accessible alternatives for the detection of Giardia and Cryptosporidium than traditional brightfield microscopes. However, the images from smartphone microscopes are noisier and require manual cyst identification by trained technicians, usually unavailable in resource-limited settings. Automatic detection of (oo)cysts using deep-learning-based object detection could offer a solution for this limitation. We evaluate the performance of four state-of-the-art object detectors to detect (oo)cysts of Giardia and Cryptosporidium on a custom dataset that includes both smartphone and brightfield microscopic images from vegetable samples. Faster RCNN, RetinaNet, You Only Look Once (YOLOv8s), and Deformable Detection Transformer (Deformable DETR) deep-learning models were employed to explore their efficacy and limitations. Our results show that while the deep-learning models perform better with the brightfield microscopy image dataset than the smartphone microscopy image dataset, the smartphone microscopy predictions are still comparable to the prediction performance of non-experts. Also, we publicly release brightfield and smartphone microscopy datasets with the benchmark results for the detection of Giardia and Cryptosporidium, independently captured on reference (or standard lab setting) and vegetable samples. Our code and dataset are available at <a href='https://github.com/naamiinepal/smartphone_microscopy'>https://github.com/naamiinepal/smartphone_microscopy</a> and <a href='https://doi.org/10.5281/zenodo.7813183'>https://doi.org/10.5281/zenodo.7813183</a>, respectively

YOLO-based object detection performance evaluation for automatic target aimbot in first-person shooter games

First-person shooter (FPS) focuses on first-person perspective action gameplay, with gunfights usually giving the player a choice of weapons, significantly impacting how the player approaches or strategies. General military-themed FPS games have realistic models with actual weapons’ shapes and characteristics. This type of game requires high aiming accuracy while using a mouse on a PC. However, not all players have a fast response time in knowing the surrounding situation. New players may need aid when targeting enemies in the FPS world. One popular yet underhanded method is injecting a program code using a dynamic-link library (DLL) to manipulate memory and asset data from the game. Instead of DLL, we promote a novel approach using the player’s real-time game screen, detecting the person without injecting program code into the game. The you only look once (YOLO) algorithm is used as an object detector model since it can process images in real time for up to 45 frames per second. The proposed object detection has an outstanding performance with 65% accuracy, 98% precision, and 61% recall of 51 tests for each game. YOLO’s fastest detection speed produces an average of 35 FPS on the YOLO tiny variant using a mixed precision (half) graphics processing unit (GPU).

A robust method for multi object tracking in autonomous ship navigation systems

A novel Maritime Multi-Object Tracking method is proposed, combining a deep learning-based object detector with target association algorithms to achieve robust sea-surface multi-object tracking. Specifically, the Multi-Object Tracking employs You Only Look Once version 7 detector for object detection. In the data association part, a module for onboard camera motion compensation is developed, a maritime dynamic spatial information-based intersection-over-union is presented as a similarity metric, and a progressive refinement cascade matching strategy is designed to enhance the tracker's sea-surface multi-object tracking capabilities. The Jari Maritime Tracking Dataset is utilised to validate the effectiveness performance of the proposed method. Experimental results demonstrate that compared to the earlier process, the proposed method exhibits a significant enhancement in multiple object tracking accuracy, with an increase of 27.8% and an achieved score of 81.3. In particular, it reduces the number of identifications switching and missed targets, achieving holistically preferable performance. Meanwhile, the speed of the proposed Multi-Object Tracking fulfils the engineering application requirements for an autonomous ship navigation system.

OffloaD: Detection Failure-based Scheduler for Offloading Object Detection

The current times ask for resource-constrained devices such as drones, light mobile robots, XR glasses, or mobile phones to perform object detection efficiently and in real time. However, when executed on the device, object detection fails to achieve the accuracy requirements. In addition, depending on a powerful edge to offload computation is time-consuming as it requires the transmission of a large amount of data and limits the usage in degraded network conditions. In this paper, we propose OffloaD, a novel offloading scheduler based on the detection of failures for object detection. The method addresses the problem of offloading between a resource-constrained device and an edge by leveraging the accuracy of the object detection model at the device and the network conditions. Accuracy is considered through a detection failure metric that consists of introspective, golden and latency scores. Our main contribution is the offloading scheduler based on the detection of failures of the object detector, which reduces unnecessary data transmission without accuracy degradation. Our method demonstrates that offloading to expensive detectors will not always increase performance and enables detection results and failure estimations even if the network is degraded or unavailable. We implement OffloaD in a physical testbed and perform extensive comparisons against several baselines. The experiments suggest that our approach improves the overall precision-latency performance by reducing the end-to-end latency between 18% and 37% in ideal network conditions and a low latency increase in degraded network conditions while obtaining an accuracy with minimum degradation compared to the baselines. In addition, comparisons to related work methods demonstrate the state-of-the-art performance of OffloaD.

VLM-guided Explicit-Implicit Complementary novel class semantic learning for few-shot object detection

Few-shot object detection (FSOD) aims at learning a novel class object detector with abundant base class samples and a limited number of novel class samples. Some recent methods assume that base class images contain unlabeled novel class instances and mine these instances to address the data scarcity of novel classes. Despite achieving competitive results, these methods face two issues: First, mined instances are sampled around the provided few-shot instances, establishing biased data distributions to represent novel classes. Second, the rich novel class semantics contained in base class instances are largely ignored. Benefits from vision-language models (VLMs), we propose a VLM-guided Explicit-Implicit Complementary novel class semantic learning (VEIC) method for FSOD. VEIC incorporates a VLM-cooperated sample mining module that leverages the powerful semantic alignment capability of VLMs to mine representative unlabeled explicit novel instances. In addition, for implicit novel instances (ambiguous base instances contain novel class semantics), we employ a spatial-wise feature re-expression module to capture prominent novel class local patches. Then a dynamic class-transition label assign strategy is proposed to learn from implicit novel instances by utilizing deep network memorization, where implicit novel instances are regarded more as novel classes during the early training and later regarded more as base classes. Experimentally, learning novel class semantics from two complementary types of instances reaches the best performance. Furthermore, learning exclusively from implicit novel instances extends the pseudo-labeling method to a practical scenario where base class images do not contain unlabeled explicit novel instances.

Robotic Grasping of Unknown Objects Based on Deep Learning-Based Feature Detection.

In recent years, the integration of deep learning into robotic grasping algorithms has led to significant advancements in this field. However, one of the challenges faced by many existing deep learning-based grasping algorithms is their reliance on extensive training data, which makes them less effective when encountering unknown objects not present in the training dataset. This paper presents a simple and effective grasping algorithm that addresses this challenge through the utilization of a deep learning-based object detector, focusing on oriented detection of key features shared among most objects, namely straight edges and corners. By integrating these features with information obtained through image segmentation, the proposed algorithm can logically deduce a grasping pose without being limited by the size of the training dataset. Experimental results on actual robotic grasping of unknown objects over 400 trials show that the proposed method can achieve a higher grasp success rate of 98.25% compared to existing methods.

EL-YOLO: An efficient and lightweight low-altitude aerial objects detector for onboard applications

Existing deep learning-based low-altitude small object detectors are typically complex in model architecture and demand substantial computational resources, making deployment for real-time detection tasks on edge computing devices challenging. To address this issue, we proposed EL-YOLO, an efficient and lightweight onboard applications object detector. Initially, in order to maintain a lightweight design and improve model accuracy, we propose a novel approach called Sparsely Connected Asymptotic Feature Pyramid Network(SCAFPN). This approach aims to eliminate inter-layer interference during feature fusion, thereby enhancing the model’s performance. Furthermore, to establish long-range contextual relationships for small object scale information, we devise a Cross-Space Learning Multi-Head Self-Attention mechanism (CSL-MHSA). To assess EL-YOLO capability for onboard small object detection, we deploy it on the embedded NVIDIA Jetson Xavier Nx platform and employ NVIDIA TensorRT FP16 quantization acceleration. On the VisDrone2019-DET and AI-TOD datasets, EL-YOLO demonstrates a 12.4% and 1.3% improvement in mAP50 compared to YOLOv5s. In comparison with the state-of-the-art YOLOv8s proposed in 2023, it achieves a respective 2.8% and 10.7% increase in mAP50. Moreover, the inference speed for the Small model reaches 24 frames per second, while the Nano model achieves 35 frames per second. This method holds promise for direct integration onto unmanned aerial vehicles for aerial small object detection tasks in the future.

LiDAR-Based 3D Temporal Object Detection via Motion-Aware LiDAR Feature Fusion.

Recently, the growing demand for autonomous driving in the industry has led to a lot of interest in 3D object detection, resulting in many excellent 3D object detection algorithms. However, most 3D object detectors focus only on a single set of LiDAR points, ignoring their potential ability to improve performance by leveraging the information provided by the consecutive set of LIDAR points. In this paper, we propose a novel 3D object detection method called temporal motion-aware 3D object detection (TM3DOD), which utilizes temporal LiDAR data. In the proposed TM3DOD method, we aggregate LiDAR voxels over time and the current BEV features by generating motion features using consecutive BEV feature maps. First, we present the temporal voxel encoder (TVE), which generates voxel representations by capturing the temporal relationships among the point sets within a voxel. Next, we design a motion-aware feature aggregation network (MFANet), which aims to enhance the current BEV feature representation by quantifying the temporal variation between two consecutive BEV feature maps. By analyzing the differences and changes in the BEV feature maps over time, MFANet captures motion information and integrates it into the current feature representation, enabling more robust and accurate detection of 3D objects. Experimental evaluations on the nuScenes benchmark dataset demonstrate that the proposed TM3DOD method achieved significant improvements in 3D detection performance compared with the baseline methods. Additionally, our method achieved comparable performance to state-of-the-art approaches.

Green pepper fruits counting based on improved DeepSort and optimized Yolov5s.

Green pepper yield estimation is crucial for establishing harvest and storage strategies. This paper proposes an automatic counting method for green pepper fruits based on object detection and multi-object tracking algorithm. Green pepper fruits have colors similar to leaves and are often occluded by each other, posing challenges for detection. Based on the YOLOv5s, the CS_YOLOv5s model is specifically designed for green pepper fruit detection. In the CS_YOLOv5s model, a Slim-Nick combined with GSConv structure is utilized in the Neck to reduce model parameters while enhancing detection speed. Additionally, the CBAM attention mechanism is integrated into the Neck to enhance the feature perception of green peppers at various locations and enhance the feature extraction capabilities of the model. According to the test results, the CS_YOLOv5s model of mAP, Precision and Recall, and Detection time of a single image are 98.96%, 95%, 97.3%, and 6.3 ms respectively. Compared to the YOLOv5s model, the Detection time of a single image is reduced by 34.4%, while Recall and mAP values are improved. Additionally, for green pepper fruit tracking, this paper combines appearance matching algorithms and track optimization algorithms from SportsTrack to optimize the DeepSort algorithm. Considering three different scenarios of tracking, the MOTA and MOTP are stable, but the ID switch is reduced by 29.41%. Based on the CS_YOLOv5s model, the counting performance before and after DeepSort optimization is compared. For green pepper counting in videos, the optimized DeepSort algorithm achieves ACP (Average Counting Precision), MAE (Mean Absolute Error), and RMSE (Root Mean Squared Error) values of 95.33%, 3.33, and 3.74, respectively. Compared to the original algorithm, ACP increases by 7.2%, while MAE and RMSE decrease by 6.67 and 6.94, respectively. Additionally, Based on the optimized DeepSort, the fruit counting results using YOLOv5s model and CS_YOLOv5s model were compared, and the results show that using the better object detector CS_YOLOv5s has better counting accuracy and robustness.

DBI-Attack: Dynamic Bi-Level Integrated Attack for Intensive Multi-Scale UAV Object Detection

Benefiting from the robust feature representation capability of convolutional neural networks (CNNs), the object detection technology of intelligent high-altitude UAV remote sensing has been developed rapidly. In this field, the adversarial examples (AEs) pose serious security risks and vulnerabilities to deep learning-based systems. Due to the limitation of object size, image degradation, and scene brightness, adding adversarial disturbances to small and dense objects is extremely challenging. To study the threat of AE for UAV object detection, a dynamic bi-level integrated attack (DBI-Attack) is proposed for intensive multi-scale UAV object detection. Firstly, we use the dynamic iterative attack (DIA) method to generate perturbation on the classification level by improving the momentum iterative fast gradient sign method (MIM). Secondly, the bi-level adversarial attack method (BAAM) is constructed to add global perturbation on the decision level for completing the white-box attack. Finally, the integrated black-box attack method (IBAM) is combined to realize the black-box mislabeling and fabrication attacks. We experiment on the real drone traffic vehicle detection datasets to better evaluate the attack effectiveness. The experimental results show that the proposed method can achieve mislabeling and fabrication attacks on the UAV object detectors in black-box conditions. Furthermore, the adversarial training is applied to improve the model robustness. This work aims to call more attention to the adversarial and defensive aspects of UAV target detection models.