Computer vision for plant and animal inventory
Metadata[+] Show full item record
The population, composition, and spatial distribution of the plants and animals in certain regions are always important data for natural resource management, conservation and farming. The traditional ways to acquire such data require human participation. The procedure of data processing by human is usually cumbersome, expensive and time-consuming. Hence the algorithms for automatic animal and plant inventory show their worth and become a hot topic. We propose a series of computer vision methods for automated plant and animal inventory, to recognize, localize, categorize, track and count different objects of interest, including vegetation, trees, fishes and livestock animals. We make use of different sensors, hardware platforms, neural network architectures and pipelines to deal with the varied properties and challenges of these objects. (1) For vegetation analysis, we propose a fast multistage method to estimate the coverage. The reference board is localized based on its edge and texture features. And then a K-means color model of the board is generated. Finally, the vegetation is segmented at pixel level using the color model. The proposed method is robust to lighting condition changes. (2) For tree counting in aerial images, we propose a novel method called density transformer, or DENT, to learn and predict the density of the trees at different positions. DENT uses an efficient multi-receptive field network to extract visual features from different positions. A transformer encoder is applied to filter and transfer useful contextual information across different spatial positions. DENT significantly outperformed the existing state-of-art CNN detectors and regressors on both the dataset built by ourselves and an existing cross-site dataset. (3) We propose a framework of fish classification system using boat cameras. The framework contains two branches. A branch extracts the contextual information from the whole image. The other branch localizes all the individual fish and normalizes their poses. The classification results from the two branches are weighted based on the clearness of the image and the familiarness of the context. Our system achieved the top 1 percent rank in the competition of The Nature Conservancy Fisheries Monitoring. (4) We also propose a video-based pig counting algorithm using an inspection robot. We adopt a novel bottom-up keypoint tracking method and a novel spatial-aware temporal response filtering method to count the pigs. The proposed approach outperformed the other methods and even human competitors in the experiments.