Animal Detectors

The autocopy.py program can be configured to run an animal detection algorithm on images as they are copied from an SD card. The Camera Trap Tools suite is currently designed to support only binary detectors, i.e., algorithms that indicate the presence or absence of the species of interest. The animal detection results are used to create a draft video segmentation indicating which frames of the time-lapse video contain animals. (See the documentation for the create_annotations.py and annotator.py programs.)

Creating Animal Detectors

Two example animal detector implementations are provided:

• The utils/genericdetector folder contains an image classifier-based algorithm that indicates if one or more animals of the species of interest is present in an image, but does not indicate where the animals are in the image. A Google Colab notebook is provided showing one way to train an image classifier for use with the generic detector.
• The utils/tortoisedetector folder contains an object detection-based algorithm that indicates if one or more gopher tortoises are present in an image and provides bounding box coordinates indicating the location of each tortoise. These bounding boxes are displayed by the annotator.py program.

The examples provided use TensorFlow, but you are free to use whatever machine learning library you wish in your implementations.

The class TrailCamObjectDetector, found in the utils/animal_detector.py file, is the abstract base class for animal detectors. To create an animal detector that works with your species of interest, you should create a new class that inherits from TrailCamObjectDetector and implements the method _object_detection(self, image_file) where image_file is the file path of the image to run the animal detector on. The result should be an N x 4 numpy array of bounding boxes indicating the extent of each of the N animals detected.

You can implement an image classifier instead of an object detector. In that case, if no animal was detected, return an empty list; if an animal was detected, return a 1 x 4 numpy array containing a single bounding box the size of the entire image.

You might also wish to implement the method _postprocessBoxes(self, boxes) where boxes is the numpy array returned by the _object_detection method. The purpose of this method is to perform any application- or camera-specific postprocessing to eliminate spurious boxes.