Adaptive Ensemble Learning for Object Counting under varying Environmental Conditions

Object counting, such as counting people in frames with varying crowd densities, remains challenging due to diverse and dynamic environmental conditions within or across image sequences.Traditional counting models typically specialize in either dense or sparse crowd scenes, limiting their effectiveness for images with mixed densities, where one region might be dense while another sparse.In this paper, an adaptive ensemble learning framework is introduced that is capable of context awareness to enable reliable object counting across varying environmental conditions.The method dynamically combines multiple specialized models optimized for different density scenarios, leveraging their individual strengths based on the spatial density distribution within each frame.Using an adaptive weighting mechanism, this framework selects the optimal combination of predictions from specialized models (for example: Faster R-CNN/Yolo for sparse scenarios and CSRNet for dense scenarios), ensuring accurate counting in complex scenes.The resulting gain in counting accuracy is especially valuable for autonomous aerial robots, which depend on dependable crowd-size estimates to support surveillance, search-and-rescue, and other safety-critical missions.Experimental results demonstrate that this adaptive ensemble learning approach significantly enhances counting accuracy compared to single-model solutions, effectively handling heterogeneous crowd distributions and environmental variability.