CONTEXT-ADAPTIVE METHOD FOR OBJECT DETECTION IN VIDEO STREAMS

The work is devoted to the development of a context-adaptive method for object detection in video streams that dynamically responds to environmental conditions. The relevance of the topic is explained by the need to increase the reliability of assistive systems for visually impaired people and other real-world applications, where variable weather and lighting conditions significantly reduce detection accuracy. The subject of the article is the study of multimodal fusion of acoustic, video, and LiDAR data for object recognition tasks. The goal of this paper is to propose and experimentally validate a method of adaptive preprocessing activation triggered by acoustic artifact classification. The task of this work is to analyze state-of-the-art preprocessing approaches (derain, defog, low-light enhancement), select appropriate acoustic classification models (e.g., PANNs, YAMNet), integrate LiDAR for spatial complementarity, and evaluate the impact of different preprocessing chains on detection metrics. Methods such as comparative analysis, experimental benchmarking of YOLO and DETR models, acoustic signal classification, and multimodal data fusion were applied. The results of the work include a confirmed increase in accuracy (mAP, Precision, Recall, IoU) and stability of detection under adverse conditions when using adaptive preprocessing pipelines, with YOLOv9m and YOLOv10m models showing the most balanced performance. Further research will focus on extending the model with full LiDAR integration, optimizing computational efficiency for mobile/embedded platforms, and scaling the approach for broader classes of environmental challenges such as fog, snow, and urban noise.