Smartphone-Based CameraTraps: On-Device Wildlife Detection : Comparing Detection Accuracy, Bandwidth Efficiency, and Per-formance Against Traditional Camera Traps

Wildlife monitoring plays a critical role in conservation efforts, enabling the tracking of species, the detection of illegal activities such as poaching, and a better understanding of animal behaviours. While traditional camera traps have proven effective, they encounter limitations, particularly regarding bandwidth, when relying on cloud-based systems for data transmission and processing. This thesis investigates the feasibility of replacing conventional digital camera traps with smartphone-based systems for wildlife monitoring. To explore this, a mobile proof-of-concept application incorporating the MegaDetector model for on-device object detection was developed. The system was supported by a custom database and web application for data handling and analysis. Field testing was conducted at Kolmården Wildlife Park, where the new smartphone-based system was compared against the Qulinda camera trap solution. Evaluation metrics included detection accuracy, long-distance performance, and bandwidth efficiency. The results show that the smartphone-based system consistently outperformed the Qulinda model, achieving higher precision (0.981 at a 50% confidence threshold), particularly in animal detection at varying distances. The system also improved bandwidth efficiency by uploading only positively identified frames, although some image redundancy persisted. Challenges encountered included class imbalance in detection data, battery limitations, and environmental factors affecting hardware performance. The study demonstrates that low-cost, mobile-based solutions can effectively support wildlife conservation and serve as a viable alternative to traditional camera traps.