A YOLOV4 METHOD FOR WILD ANIMAL COUNTING AND BEHAVIOUR DETECTION USING SMALL-SIZED CAMERA-TRAP IMAGE DATASET

Many deep learning-based solutions proposed to automate the analysis of camera-trap images for animal counting and behaviour detection deployed image classifiers that produce image-level labels, tackle animal counting as a classification task, and use images with one animal or attribute. They also used large image datasets which are costly, time-consuming and laborious to collect and annotate, not feasible for rare/elusive species and resource-constrained projects and did not explain the generalization of the models on images with untrained image backgrounds. This study developed animal counting and behaviour detection models for wild animals using You-Only- Look-Once (YOLO) and small-sized datasets with 20,110 camera-trap images. The study results show that using appropriate technologies including transfer learning, data augmentation and efficient data splitting methods, it is feasible to develop high-accurate and location- invariant object detection models using small-sized image datasets. Despite high performance, the animal counting model did not perform well on crowding/interacting animals including Guineafowl, Elephant, Lion, Zebra, Wildebeest, Baboon and Giraffe and it misclassified a significant number of Wildebeest as Buffalo and Zebra, but few Buffalo and Zebra were misclassified as Wildebeest. The behaviour detection model performed well on all behaviours except interacting. Model’s poor performance on interacting is ascribed to its messy and small training set compared to other behaviour classes. The selection of an optimal confidence threshold appropriate for a particular dataset and increasing data diversity of the training set can significantly improve recall while reducing false positives.