Object recognition models trained on static images often perform suboptimally when applied directly to video sequences. This is due to factors such as redundant computations across consecutive frames, lack of temporal consistency in recognition results, and a higher incidence of false positives. In this work, we propose a strategy to adapt these models for more efficient and robust use in dynamic video contexts. Our approach is based on temporal processing techniques that reduce the frequency of full-model inferences by validating object presence through temporal recurrence analysis over short frame windows. This helps to maintain recognition accuracy while reducing computational demands. To further improve temporal coherence, we incorporate filtering mechanisms and consensus validation across frames, which enhance stability and reduce spurious detections or isolated false positives. Additionally, we explore preprocessing techniques to improve input quality, such as correcting fisheye distortion commonly found in IP camera lenses, which can degrade object appearance near image boundaries. The proposed method is lightweight and modular, allowing it to extend existing image-based recognition systems to video applications without requiring retraining or architectural changes. Experimental results in video streams show improvements in both computational performance and recognition reliability. Our findings suggest that taking advantage of temporal redundancy and visual consistency can significantly enhance the applicability of image-based object recognition models in real-world video environments.