A CASCADE METHOD FOR IMPROVING OBJECT DETECTION SYSTEMS

Abstract

The paper addresses the problem of instability of neural network-based object detection systems deployed on unmanned aerial vehicles (UAVs) under dynamically changing environmental conditions. It is shown that conventional universal models, including the YOLO family architectures, demonstrate a significant degradation of key performance metrics such as Precision and Recall when operating under varying weather conditions, illumination changes, and different object scales. The aim of the study is to improve the efficiency and robustness of real-time object detection systems by developing an adaptive cascade-based processing method. The proposed Cascade YOLO approach is based on sequential utilization of specialized neural network models combined with an adaptive switching mechanism depending on the confidence level of the current detection. Unlike classical ensemble approaches, the proposed method avoids reprocessing the same frame and instead applies the next model to a new incoming frame, which prevents latency accumulation and preserves real-time performance. The research methodology includes the development of a mathematical model of the cascade algorithm, its software implementation, and experimental validation on heterogeneous video datasets with varying observation conditions. Comparative analysis with a baseline Single-YOLO model demonstrated an increase in Precision up to 0.867 and Recall up t 0.824, with only a minor decrease in processing speed (within 5–10 %). The obtained results confirm the effectiveness of the proposed approach in enhancing the reliability of UAV-based computer vision systems. The practical significance of the research lies in the possibility of integrating the method into onboard data processing systems to ensure stable performance under uncertainty and dynamic environmental conditions.