Sugarcane seedling emergence is important for sugar production. Manual counting is time-consuming and hardly practicable for large-scale field planting. Unmanned aerial vehicles (UAVs) with fast acquisition speed and wide coverage are becoming increasingly popular in precision agriculture. We provide a method based on improved Faster RCNN for automatically detecting and counting sugarcane seedlings using aerial photography. The Sugarcane-Detector (SGN-D) uses ResNet 50 for feature extraction to produce high-resolution feature expressions and provides an attention method (SN-block) to focus the network on learning seedling feature channels. FPN aggregates multi-level features to tackle multi-scale problems, while optimizing anchor boxes for sugarcane size and quantity. To evaluate the efficacy and viability of the proposed technology, 238 images of sugarcane seedlings were taken from the air with an unmanned aerial vehicle. Outcoming with an average accuracy of 93.67%, our proposed method outperforms other commonly used detection models, including the original Faster R-CNN, SSD, and YOLO. In order to eliminate the error caused by repeated counting, we further propose a seedlings de-duplication algorithm. The highest counting accuracy reached 96.83%, whilst the mean absolute error (MAE) reached 4.6 when intersection of union (IoU) was 0.15. In addition, a software system was developed for the automatic identification and counting of cane seedlings. This work can provide accurate seedling data, thus can support farmers making proper cultivation management decision.

Dragon fruit (Hylocereus undatus) is a tropical and subtropical fruit that undergoes multiple ripening cycles throughout the year. Accurate monitoring of the flower and fruit quantities at various stages is crucial for growers to estimate yields, plan orders, and implement effective management strategies. However, traditional manual counting methods are labor-intensive and inefficient. Deep learning techniques have proven effective for object recognition tasks but limited research has been conducted on dragon fruit due to its unique stem morphology and the coexistence of flowers and fruits. Additionally, the challenge lies in developing a lightweight recognition and tracking model that can be seamlessly integrated into mobile platforms, enabling on-site quantity counting. In this study, a video stream inspection method was proposed to classify and count dragon fruit flowers, immature fruits (green fruits), and mature fruits (red fruits) in a dragon fruit plantation. The approach involves three key steps: (1) utilizing the YOLOv5 network for the identification of different dragon fruit categories, (2) employing the improved ByteTrack object tracking algorithm to assign unique IDs to each target and track their movement, and (3) defining a region of interest area for precise classification and counting of dragon fruit across categories. Experimental results demonstrate recognition accuracies of 94.1%, 94.8%, and 96.1% for dragon fruit flowers, green fruits, and red fruits, respectively, with an overall average recognition accuracy of 95.0%. Furthermore, the counting accuracy for each category is measured at 97.68%, 93.97%, and 91.89%, respectively. The proposed method achieves a counting speed of 56 frames per second on a 1080ti GPU. The findings establish the efficacy and practicality of this method for accurate counting of dragon fruit or other fruit varieties.

Underwater multi-target tracking is one of the key technologies for military missions, including patrol and combat in the crucial area. Since the underwater environment is complex and targets’ trajectories may intersect when they are in a dense area, it is challenging to guarantee the precision of observed information. In order to provide high-precision underwater localization and tracking services over an underwater monitoring network, a dynamic network resource allocation mechanism and an underwater multi-target tracking algorithm based on a two-layer particle filter with distributed probability fusion (TLPF-DPF) are proposed. The position estimation model based on geometric constraints and the dynamic allocation mechanism of network resources based on prior position estimation are designed. Using the improved filtering algorithm with known initial states, the reliable tracking of multiple targets with trajectory intersection in a small area under complex noises is achieved. In the non-Gaussian environment, the average positioning error of TLPF-DPF is less by nearly 30% than alternative algorithms. When switching from a Gaussian environment to a non-Gaussian environment, the performance degradation of TLPF-DPF is less than 12%, which exhibits stability compared with other algorithms when targets are close to each other with crossing trajectories.

Reinforcement learning (RL) has been shown to be effective in path planning. However, it usually requires exploring a sufficient number of state–action pairs, some of which may be unsafe when deployed in practical obstacle environments. To this end, this paper proposes an end-to-end planning method based model-free RL framework with optimization, which can achieve better learning performance with a safety guarantee. Firstly, for second-order drone systems, a differentiable high-order control barrier function (HOCBF) is introduced to ensure the output of the planning algorithm falls in a safe range. Then, a safety layer based on the HOCBF is proposed, which projects RL actions into a feasible solution set to guarantee safe exploration. Finally, we conducted a simulation for drone obstacle avoidance and validated the proposed method in the simulation environment. The experimental results demonstrate a significant enhancement over the baseline approach. Specifically, the proposed method achieved a substantial reduction in the average cumulative number of collisions per drone during training compared to the baseline. Additionally, in the testing phase, the proposed method realized a 43% improvement in the task success rate relative to the MADDPG.