Deep Learning for Detecting Abandoned Dogs

This research paper presents a methodology consisting of an algorithm and a workflow for finding abandoned dogs in natural surroundings. We propose a temporal-contextual methodology for identifying abandoned pets in public areas. This involves employing a temporal rule alongside context-sensitive object identification, whereby dog bounding boxes are deliberately expanded to encompass proximate visual indicators suggestive of abandonment. The proposed approach uses object detection techniques, trajectory analysis, and image segmentation to quickly differentiate between abandoned and owned dogs. The research addresses key challenges, such as data scarcity and the complexity of distinguishing between abandoned and accompanied dogs. To address the issue of sufficient and adequate training corpus to identify an abandoned dog, the research article employs single-channel image augmentation methods that improve model recall and precision by 4%. Several object detection algorithms were evaluated, and our findings indicate that single-stage detectors like YOLOv8 achieved a better trade-off between classification performance and speed for detecting abandoned dogs compared to multi-stage detectors like Faster R-CNN, reaching a mean average precision up to 86% and an inference time of 0.3 ms per frame. This research study contributes to animal welfare, biodiversity conservation, and public safety by providing a scalable solution for monitoring abandoned animals in diverse environments. The findings demonstrate the impact of object detection techniques on improving the generalization of deep learning models for real-world applications.