Camera traps are a critical tool in wildlife and transportation ecology, used to measure occupancy, movement, activity, and basic demographic characteristics. Ancillary motion in the image scene, such as moving vegetation, can cause the camera to trigger, leading to collections of "no-animal" images, which can contribute to unnecessary effort by analysts in proportion to the number of empty images. A significant advantage of image datasets collected from a fixed location, is that historical conditions in the imagery can be used to approximate an expected background condition for each image.
We developed an automated series of processes to detect images with and without objects of interest within a web-system used to manage camera trap images (https://wildlifeobserver.net). The web-system is available for anyone to use to manage camera trap projects and is primarily oriented around monitoring of wildlife adjacent to transportation systems. We describe a workflow for automatically detecting images in a dataset of camera trap imagery that do or do not include an animal. First, imagery is segregated based on image characteristics. Then, imagery is segmented using a Gaussian mixing approach. When an image containing an animal is compared against the multimodal distribution of expected pixel values, strong differences emerge at the location of the animal.
We used a series of test settings with minimized background complexity: wildlife underpasses and culverts. We achieved a 94% true negative rate (images classified as empty, and which did not contain an animal) using a classification scheme that incorporates image timestamp, median image approximation, and a convolutional neural network animal identifier (MLWIC). The accuracy of MLWIC decreased with background complexity. Images classified as empty were automatically tagged within the web-system, whereas those with animals were provisionally tagged with the animals' species or group (e.g., "deer"), if identifiable. We also tested two available systems that theoretically identify organisms in images: 1) MLWIC, mentioned above - a protocol in R that was reported in the literature to have accuracies as high as 98% for imagery from within the training dataset, and 83% for out-of-sample imagery. When tested with our camera trap imagery, however, MLWIC achieved 14.4% top-1 guess accuracy and 28.8% top-3 guess accuracy for animal-species. 2) SEEK is an app available from iNaturalist that can identify plant, animal and fungal organisms to the genus or species level. We found that the app could not identify species that were moving or far away from the camera, but still identifiable by a wildlife technician. Furthermore, SEEK did not differentiate between "no animal" and "unable to identify animal". We conclude that our tool can remove the majority of images that contain no animal, but that there are still errors in classification that could result in removal of images that actually do contain an animal.