This paper presents a hybrid approach with machine learning and light-weight radio propagation model to estimate wireless communication range between Unmanned Aerial Vehicles (UAVs) flying at low altitude and their operators for crop-spraying in orchards. Additionally, it clarifies key factors to realize the approach by comparing measured data in a real field and simulation results. Simulation and measurement results shown in this paper reveals significant discrepancies between Irregular Terrain Model (ITM) and raytracing simulation results and some different patterns of signal attenuation due to vegetation and artificial obstacles such as trees and buildings between UAVs and their operators. The authors show that 3D special information, which is voxelized point cloud data measured in a real field, can improve the accuracy of results of the light-weight radio propagation model and discuss on detailed design of our approach with consideration of radio propagation characteristics.
This paper presents ChickOut, an intelligent poultry retail and inventory system that eliminates the need for barcodes by integrating deep learning-based object recognition and automated weighing for real-time checkout and inventory management. The system was developed to address inefficiencies in traditional retail operations involving non-barcoded products such as fresh chicken meat, which typically require manual logging. A custom dataset of 4,207 images (6,638 annotations) across five chicken parts—breast, leg, thigh, wing, and quarter—was used to train a YOLOv7 detection model. After 300 training epochs, the model achieved an mAP@0.5 of 97.8 %, $\text{mAP} {@} 0.5: 0.95$ of 91.2 %, precision of 96.5 %, and recall of 95.8 %, enabling reliable real-time detection under varying lighting and orientation. The weighing module, composed of a load cell and HX711 amplifier interfaced with an Arduino Uno, provides $\pm 1 ~\mathrm{g}$ accuracy and an average response time of 0.3 seconds. The Django-based backend automatically computes prices, updates stock in real time, and issues low-stock alerts, achieving 100 % synchronization accuracy between sensor data and database records. An integrated web-based analytics dashboard visualizes sales trends, top-selling items, and vendor performance with $<2$-second data refresh latency for multiple terminals. By removing the reliance on barcodes and combining AI vision, embedded sensing, and analytics, ChickOut enhances operational efficiency, transparency, and sustainability in poultry retail—advancing UN SDGs 9 and 12.
House object Accurate localization and classification of dental structures in the radiographic images are crucial for both automatic dental diagnosis and treatment planning. This paper proposes a deep learning-based tooth detection framework that integrates a multi-stage architecture, comprising a backbone for hierarchical feature extraction, a neck module for multi-scale feature refinement, and a detection head for bounding box regression and class prediction. The backbone employs a series of convolution-batch normalization-activation blocks to capture low- and mid-level anatomical patterns from input dental images. The neck module fuses multi-scale features using upsampling and concatenation operations, enhancing contextual representation. Then, the final predictions, including tooth class probabilities, objectness scores, and optimized bounding boxes, are generated in the detection head. Experimental evaluation on multiple tooth categories reveals powerful detection performance, with an overall precision of 0.786, an overall recall of 0.848, mAP@50 of 0.881, and mAP@50-95 of 0.656. Class-wise results display excellent performances for Central Incisors, Canines, and Lateral Incisors, having mAP@50-95 values of $0.845,0.808$, and 0.823, respectively, while comparatively low scores have been observed for some classes like the 2nd Molar. In general, the proposed model effectively balances accuracy and robustness across diverse tooth types.
Surround-view monitoring systems for heavy-duty vehicles must maintain reliable BEV (bird's-eye-view) imagery over long-term operation in harsh environments. Classical geometric calibration with checkerboards or ChArUco boards provides accurate intrinsics and extrinsics at installation time, but cannot easily cope with parameter drift caused by vibration, thermal cycles, and minor impacts. This paper proposes a twostage AVM (around-view monitoring) framework that combines pattern-based commissioning with learning-based maintenance of camera intrinsics. A learning-based calibration network predicts a ray-direction field from natural driving scenes and fits a parametric fisheye model, enabling both user-initiated recalibration and automatic monitoring of intrinsics drift during runtime. Experiments on a four-camera surround-view platform show that the proposed method reduces reprojection error from 102.44/107.52 pixels (GeoCalib ${ }_{\text{radial/gen }}$) to 47.304 pixels (Ours ${ }_{\text{dist }}$), lowers angular error from $4.28 / 7.24^{\circ}$ to $0^{\circ}. 77^{\circ}$, and decreases vertical/horizontal field-of-view errors from $7.45 / 14.88^{\circ}$ to $0.49 / 1.26^{\circ}$, while matching or improving the accuracy of recent learning-based calibration. These gains translate into visibly more stable surround-view images compared to purely geometric baselines.
