[en] In African open protected areas, large mammals are often surveyed using manned aircrafts which actively count the animals in sample strips for later density extrapolation to the whole area. Nevertheless, this method may be biased among others by the observer’s detection capability. The use of on-board oblique cameras has recently shown an increase in counting accuracy as a result of indirect photo-interpretation. While this approach appears to reduce some biases, the processing time of the generated data is currently a bottleneck. In recent years, Deep Learning (DL) techniques through dense convolutional neural networks (CNNs) have emerged as a very promising avenue for managing such datasets. However, we are not yet at the stage of full automation of the process (i.e. from acquisition to population estimation). Three challenges were identified: 1) reducing false positives, 2) increasing the precision in close-by individuals, and 3) properly managing the overlap between images to avoid double counting. We focused on the two first aspects and developed a new point-based DL model inspired by crowd counting, that was applied on a challenging oblique aerial dataset containing free ranging livestock herds in heterogeneous open arid landscapes. The model’s performances were then evaluated using localization and counting metrics. The DL model achieved a global F1 score of 0.74 and a RMSE of 9.8 animals per 24 megapixel image, at a processing speed of 3.6 s/image. It showed a valuable ability to detect both isolated animals and those in dense herds. This is auspicious for automation of African mammal surveys but the developed approach still needs to be improved to manage double counting on entire transects. These results emphasize the importance of standardization of data acquisition, with strong spatial and temporal heterogeneities, in order to build robust models that can be used in similar environments and conditions.
