23 October 2020 in IEEE Access
Vehicle detection in aerial videos often requires post-processing to eliminate false detections. In this paper, we establish a spatial-temporal processing scheme to improve the performance of automatic vehicle detection using wide-area aerial video datasets. We replace the thresholding step of existing detection algorithms with multi-neighborhood hysteresis thresholding for foreground pixel classification. This is followed by spatial post-processing, which includes morphological opening and closing to shape and prune the detected objects. We then apply temporal post-processing to further reduce false detections. Experimental results demonstrate that the proposed thresholding and spatio-temporal processing improve vehicle detection performance. We compute the average F-score and percentage of wrong classifications on a few detection algorithms with and without the proposed scheme by using registered frames from five aerial video datasets and two public online sample datasets for verification. Experimental results confirm the accuracy and efficiency of our approach, which achieves a 90% reduction in false detections and a percentage of wrong classifications below 50%.
16 June 2026 in Journal of Laser Applications
In the precision-driven domain of laser welding, achieving consistent high-quality welds while maintaining process efficiency poses a significant challenge. This paper introduces a low-latency AI-based framework designed to optimize the laser welding process in real time, addressing this challenge through advanced monitoring and control. The system comprises two key components: an AI-based monitoring system and an AI-based controller. Utilizing data from a high-speed camera and a high-speed microphone, the monitoring system integrates multiple AI models and data fusion to predict four critical quality measures. These include surface quality parameters and subsurface characteristics, such as weld depth and bonding width, with weld depth predictions achieving less than 10% error. The control system leverages these real-time assessments to compute optimal process parameters, ensuring continuous process optimization. It manages seven control parameters, including x–y beam oscillation and power modulation, with monitoring inference in approximately 2–3 ms, control computation in approximately 1 ms, and a total reaction time of approximately 10–25 ms. A stochastic digital twin simulator facilitated system testing and tuning prior to deployment. The novelty of this work lies in achieving low-latency operation under high complexity (four quality measures, seven control parameters). In surrogate-based stochastic closed-loop simulations calibrated from copper-electrode welding experiments, the proposed controller reduced the simulated quality deviation from target values by 29%–48%, depending on the quality measure.