Comparative evaluation of artificial intelligence-integrated and conventional camera traps for small mammal monitoring

Context Conventional camera traps generate vast amounts of non-target images that makes edge artificial intelligence (AI) cameras a promising solution for real-time species detection. However, the real-world effectiveness of these systems depends heavily on how models generalise from training to unpredictable field conditions. Aims We test whether within-distribution validation accuracy predicts field deployment performance for edge-deployed object detection models by examining how training-dataset background composition affects field generalisation for small mammal detection, and compare AI-integrated and conventional camera-trap detection outcomes under matched field conditions. Methods We conducted a field-based comparison between an AI-integrated camera system (MaixCAM-Pro running YOLOv11n) and a conventional trail camera (Hawkray HC-900A) for small mammal monitoring. Three object-detection models were evaluated based on training datasets with different background compositions: heterogeneous varied backgrounds (Model A), uniform white backgrounds (Model B) and a combined dataset integrating both (Model C). Key results Although Model B achieved the highest validation accuracy during training, field deployment revealed a severe performance gap. Model C demonstrated the strongest real-world performance, with 88.9% of MaixCAM-Pro captures correctly containing squirrels and a false-positive rate of 11.1%, markedly lower than Model A’s 98.4%. Conversely, Model A performed poorly in the field due to extensive background-related misclassifications. When the AI-integrated camera is equipped with Model C, it significantly reduces false triggers compared to the conventional camera while maintaining reliable small mammal (squirrel) detection. Conclusions Training-dataset background composition substantially affects field detection performance. The within-distribution validation accuracy does not reliably predict field deployment success. Balanced datasets incorporating both controlled and varied backgrounds are essential to improve model generalisation. Implications Wildlife AI applications must validate models under field conditions matching the deployment context rather than relying solely on training metrics. Furthermore, the model’s precision profile must align with the downstream ecological inference framework (e.g. occupancy modelling) to ensure data integrity.