Remote detection of humans or large animals in natural areas is an important tool for biologists and sociologists who seek to understand the interactions between humans and fauna in sensitive areas. The applications relate as well to the behavioral study of human populations (hikers, mountain bikers, skiers, hunters, forest professionals ...) as animal populations (wolves, deer, wild boars, fauna carrying ticks transmitting Lyme disease ...). The goal of the WildCount project is the development of an inexpensive hunting camera for autonomous recognition and counting of the presence of humans (anonymously) and animals into the wild and protected areas such as authorized paths of natural parks, near mountain refuges, avalanche zones ... This sensor will use inexpensive off-the-shelves components but nevertheless offering good performance (image resolution, energy consumption, etc.). The technologies envisaged will be based on deep learning techniques (i.e. neural networks). Neural networks will be trained from the human, animal and vehicle image databases depending on the installation / observation context. The calibration will be done in partnership with the technical teams of the Zone Ateliers Alpes (including the Station Alpine Joseph Fourier and the Parc National des Ecrins). Two proofs of concept were carried out in 2020 (MAIx Sipeed board and a LoRaWAN modem) and 2021 (Greenwaves GAP8POC boards). The new version of the WildCount edge sensor will be built around the Sony Spresense board, thermal cameras, visible-light cameras, PIR motion sensors, environmental sensors (temperature, humidity, pressure, etc.) and a low-power and long-range LoRaWAN communication modem (RN2483 and Lambda80) on ISM bands (868 MHz, 2.4 GHz). This hunting camera respects the privacy of any person passing in the detection field since no image will be stored or transmitted by the sensor firmware. This hunting camera only transmits occupancy counters for the different types of species recognized by the neural network. It can send alerts on detection according to a calendar loaded on the sensor. The software architecture of the project follows the principles of edge-computing architectures: the hunting camera embeds a large part of processing (image recognition) to transmit only synthetic information to the cloud via a long range and robust communication but very low throughput network (approximately 300 baud over 1% of duty cycle). The summarized data is then stored into a datacenter for triggering alerts and displaying dashboards.