At present, the protection of birds, especially endangered birds, faces major challenges. In the process of protection, birds are often mixed with various drones, and it is difficult to accurately count the number of endangered birds, which brings great difficulties to bird protection work. So tracking and identifying birds and drones is crucial. To solve these problems, this paper proposes a new multi-target tracking (MOT) model based on the combination of YOLOv9 detection algorithm and DeepSORT tracking algorithm. Firstly, the original RepNSCPELAN4 module is replaced by CAM context feature enhancement module in Backbone to improve the model’s ability to extract small target features. Following this, the AFF channel attention mechanism has been integrated with RepNSCPELAN4 in the Head section to create the Repnscpelan4-AFF module, which aims to better address semantic and scale inconsistencies. Finally, a new RepNSCPELAN4-Akconv module has been developed using the AKConv dynamic Convolution module to replace the RepNSCPELAN4 module in the original Head section, enabling the model to more effectively capture detailed and contextual information. On the Bird-Drone visible light comprehensive data set proposed in this study, the improved YOLOv9-DeepSORT model has a mAP0.50 of 81.3% for all categories and 89.1% for individual birds. Compared to the baseline YOLOv9 original model, improvements of 7.9% and 23.9% respectively. On infrared datasets, compared to the original model, the mAP0.50 of the improved model is improved by 3.2% in all categories. The accuracy of identifying individual birds and similarly shaped fixed-wing drones also improved by 2.2% and 7.5% respectively. Moreover, on the mixed visible light and infrared data sets, the model get mAP0.50 of 81.8% higher 0.9% than that of the YOLOv9. These experiments demonstrate the improved YOLOv9-DeepSORT method can expand the multiscene application range of bird recognition and tracking models, effectively promoting the extraction of video frame features in multi-target tracking.

Automated fruit-picking equipment has the potential to significantly enhance the efficiency of picking. Accurate detection and localization of fruits are particularly crucial in this regard. However, current methods rely on expensive tools such as depth cameras and LiDAR. This study proposes a low-cost method based on monocular images to achieve target detection and depth estimation. To improve the detection accuracy of targets, especially small targets, an advanced YOLOv8s detection algorithm is introduced. This approach utilizes the BiFormer block, an attention mechanism for dynamic query-aware sparsity, as the backbone feature extractor. It also adds a small-target-detection layer in the Neck and employs EIoU Loss as the loss function. Furthermore, a fused depth estimation method is proposed, which incorporates high-resolution, low-resolution, and local high-frequency depth estimation to obtain depth information with both high-frequency details and low-frequency structure. Finally, the spatial 3D coordinates of the fruit are obtained by fusing the planar coordinates and depth information. The experimental results with citrus as the target result in an improved YOLOv8s network mAP of 88.45% and a recognition accuracy of 94.7%. The recognition of citrus in a natural environment was improved by 2.7% compared to the original model. In the detection range of 30 cm~60 cm, the depth-estimation results (MAE, RSME) are 0.53 and 0.53. In the illumination intensity range of 1000 lx to 5000 lx, the average depth estimation results (MAE, RSME) are 0.49 and 0.64. In the simulated fruit-picking scenario, the success rates of grasping at 30 cm and 45 cm were 80.6% and 85.1%, respectively. The method has the advantage of high-resolution depth estimation without constraints of camera parameters and fruit size that monocular geometric and binocular localization do not have, providing a feasible and low-cost localization method for fruit automation equipment.

Aiming at the difficulty of feature extraction in complex environments during mulberry detection and the need for embedded devices to lighten the model, this study carries out lightweight improvements on the basis of the YOLOv8n model. First, the CSPPC module incorporates lightweight partial convolution (PConv) within its bottleneck structure, replacing the C2f module to enhance feature extraction efficiency. Secondly, the ADown module is used to replace the traditional downsampling module and the P-Head module is used to replace the traditional convolutional detector head with the partial convolutional detector head. Finally, a knowledge distillation technique is used to compensate for the loss of accuracy due to parameter reduction. Ablation experiments are conducted to evaluate the impact of each module on the model’s performance. The experimental results show that the improved YOLOv8 model has a precision of 88.9%, a recall of 78.1%, and an average precision of 86.8%. The number of parameters is 1.29 × 10⁶, the model size is 2.6 MB, the floating-point computation is 2.6 GFLOPs, and the frame rate reaches 19.84 FPS on the edge end. Therefore, this model provides theoretical and technical support for the deployment and application of mobile detection devices, such as automatic mulberry harvesting in practical scenarios.

As the application of Unmanned Aerial Vehicles(UAVs) becomes increasingly widespread, the identification of UAVs is of great significance in the field of security. The research of advanced identification technology can effectively deal with the illegal invasion of UAVs and reduce the threat to aviation safety. However, during the recognition process, the effectiveness of UAVs identification is often compromised in long-distance and complex environments, particularly in night-time scenarios, where accurately and reliably identifying UAVs remains a significant challenge. To overcome this, this paper proposes an improved algorithm named YOLOv9-CAG, which is based on data from multiple sensors. This algorithm integrates the detection capabilities of visible light, infrared, and audio signals. The improvements primarily encompass three key aspects: The RepNSCPELAN4 module at the end of the trunk has been replaced with a CAM context feature enhancement module to bolster the capability of extracting features from small target UAVs; A GAM attention mechanism has been integrated into the head network to enhance the model’s focus on specific areas or features of UAVs; An enhanced AKConv dynamic convolution has been implemented at the end of the head, building upon the original RepNSCPELAN4 module to more effectively capture contour details. On the Bird-UAV visible light data set, the mAP0.50 of UAVs by the improved YOLOV9-CAG model is 92.0%, which is 10.8% higher than that of the original YOLOv9 model. In terms of infrared data set, the mAP0.50 and recall of the enhanced model on the UAVs reached 86.5% and 89.2% respectively, which were also increased by 12.4% and 11.4% compared with the original YOLOv9 model, which expanded the effectiveness of the model in the infrared scene. On the audio spectrum dataset, the enhanced model demonstrated improvements in UAV recognition compared to the original YOLOv9 model, achieving increases of 8.4% in mAP0.50 and 14.3% in recall respectively. At the same time, the enhanced model in this study also has a good recognition effect on birds, achieving mAP0.50 of 85% and 94.8% under visible light and infrared conditions respectively, which is 19.8% and 1.1% higher than the original YOLOv9 model. In validation on real-world visible-light and infrared videos, the YOLOv9-CAG model demonstrated an overall average accuracy improvement of 6.8% and 3.8%, respectively, over the original YOLOv9 model. The results show that the improved YOLOv9-CAG model has excellent performance in UAVs recognition in multiple scenarios. This work pioneers a multimodal UAVs detection framework that significantly improves identification accuracy in challenging conditions, pushing the boundaries of UAVs identification technology.

The integration of Multi-layer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), and attention mechanisms has been demonstrated to significantly enhance model performance across various computer vision tasks. In this paper, a novel lightweight neural network architecture, FasterMLP, is proposed to achieve high computational efficiency and accuracy, particularly in resource-constrained and real-time applications. FasterMLP is designed to combine the local connectivity and weight-sharing properties of CNNs with the global feature representation capabilities of MLPs, while feature extraction is enhanced through the Convolutional Block Attention Module and spatial dimensions are effectively reduced using Haar wavelet downsampling without sacrificing critical feature information. The architecture, structured into four stages, has been rigorously evaluated on multiple benchmarks. On the ImageNet-1K dataset, a top-1 accuracy 3.9% higher than that of MobileViT-XXS is achieved by FasterMLP-S, while being 2\(\times\) and 2.7\(\times\) faster on GPU and CPU, respectively. On the COCO dataset, the performance of FasterMLP-L is shown to be comparable to FasterNet-L with significantly fewer parameters, and on the Cityscapes dataset, a mean Intersection-over-Union of 81.7% is achieved, surpassing existing methods such as CCNet and DANet. These results demonstrate that FasterMLP can effectively balance computational efficiency and accuracy, making it particularly suitable for visual perception tasks in resource-constrained and real-time environments such as autonomous driving. Code is available at https://github.com/windisl/FasterMLP.

Accurate drone and bird identification is vital for airspace safety and ecological protection. However, their similar appearance and behavior at long distances and small scales, combined with challenges like lighting, weather, and target size, limit single-sensor (visible or infrared) recognition performance in complex scenarios. To address these challenges, this paper proposes YOLOv10-PRD, an enhanced recognition method based on YOLOv10 with multi-source data fusion (visible, infrared, radar, and audio). By leveraging complementary information and enhancing feature extraction, the model achieves higher robustness and accuracy. The improvements include the integration of the PSContextAggregation module to more effectively capture features of distant or blurry targets. To enhance the model’s focus on critical features in complex scenarios, RFAConv is integrated at the tail end of the head network, replacing standard convolutions. Additionally, DWConv effectively preserves local features within each channel while reducing model complexity and inference time. The YOLOv10-PRD model was evaluated on multi-modal datasets including visible light, infrared, radar droppler, audio spectrum, optical fusion, and COCO. It achieved a mAP0.50 of 82.5% on the visible dataset (6.9% improvement over YOLOv10), 84.4% on infrared (2.6% improvement), 66.3% on radar droppler (1.9% improvement), and an 11.1% gain on the audio spectrum dataset. It also significantly outperformed YOLOv11, Gold-YOLO, EfficientDet, DAMO YOLO, and other mainstream models on COCO and optical mixed datasets. In video detection of small and medium targets such as birds and drones, it improved mAP0.50 by 9% and 4.4%, respectively. Deployed on low-power devices, the model maintained 41 FPS with high detection precision, meeting real-time requirements for medium- and long-range scenarios. Overall, YOLOv10-PRD balances mAP, efficiency, and adaptability, showing strong potential for practical deployment.