An enhancement of CNN applied in multi object tracking system in heterogenous zoological environment

ABSTRACT: The enhancement of the CNN algorithm successfully achieved the three specific objectives set for this study. First, the transition from binary to multi-class species classification significantly improved the model’s capability, with the classification accuracy increasing from 76.4%, and even reaching 100% in controlled evaluation scenarios compared to the original 66.7%. Second, the refinement of the CNN architecture through the integration of EfficientNetB3 resulted in a 33.3% improvement in accuracy, demonstrating superior feature extraction and generalization capabilities over the original basic CNN model, despite a slight increase in inference time. Third, the integration of the enhanced CNN with YOLOv8 for object detection and Deep SORT for object tracking provided real-time detection and tracking capabilities, achieving a mean Average Precision (mAP) of 91.5% and reducing ID switches by 78.26% compared to the baseline MOTHe configuration. These enhancements effectively addressed the limitations of the original algorithm, resulting in substantially higher classification precision, more reliable tracking under complex environmental conditions, and a scalable, modular framework capable of future upgrades. The findings validate that targeted algorithmic improvements in classification, feature extraction, and real-time tracking significantly enhance the

Filipiniana