New Scientist (01/10/13) Hal Hodson
Harvard University researcher Walter Scheirer has developed a smartphone-based machine-vision system that automatically recognizes and counts specific animals. Scheirer says the system could help biologists make quicker, more accurate judgments about the health of fragile ecosystems. Although automated camera traps are already in use, they are not selective enough. "Right now, we have to manually go through every photo to identify species and separate photos of interest from false photos," says Princeton researcher Siva Sundaresan. The system starts by scanning the environment for objects that could be the animals it is looking for. It looks for pixel clumps that are new to the scene, then studies them to determine whether they represent any of the animals it has been trained to recognize. The algorithms analyze the content of each frame and look for patterns of pixels that identify the animal. Testing has shown the system can distinguish between three different species of ground squirrel 78 percent of the time. Scheirer says the goal is to develop an inexpensive, easy-to-use system that can automatically detect animals in any environment.
Harvard University researcher Walter Scheirer has developed a smartphone-based machine-vision system that automatically recognizes and counts specific animals. Scheirer says the system could help biologists make quicker, more accurate judgments about the health of fragile ecosystems. Although automated camera traps are already in use, they are not selective enough. "Right now, we have to manually go through every photo to identify species and separate photos of interest from false photos," says Princeton researcher Siva Sundaresan. The system starts by scanning the environment for objects that could be the animals it is looking for. It looks for pixel clumps that are new to the scene, then studies them to determine whether they represent any of the animals it has been trained to recognize. The algorithms analyze the content of each frame and look for patterns of pixels that identify the animal. Testing has shown the system can distinguish between three different species of ground squirrel 78 percent of the time. Scheirer says the goal is to develop an inexpensive, easy-to-use system that can automatically detect animals in any environment.
