Embedded smart vision system based on RBF neural processor: application to human pose classification.
Nowadays, person detection is more and more necessary in the context of security applications. In this context, several systems have been proposed like single standalone video systems or PC-based solutions. If these systems can reach relative good
performances in terms of success of detection, they require however presence of human operator assistance and in general, automatic detection in these systems is not embedded because used image processing algorithms require a lot of calculations which are difficult to implement on hardware.
One solution to improve performances of such systems is to use artificial neural network whose characteristic is to emulate the reasoning of the human brain which can recognize with surprising ease, and in real tim, a pattern from the moment it was learned.
Moreover, neural network is inherently a group of elements with the same behavior, it is a candidate for a parallel architecture.
In this field, the CogniMem chip is a fully parallel silicon neural network: it is a chain of 1024 identical elements (i.e. neurons) addressed in parallel and which have their own “genetic” material to learn and recall patterns without running a single line of code and without synchronizing to any supervising unit. A resulting achievement of this architecture is a constant learning and recognition time regardless of the number of connected neurons.
Using the CogniMem chip connected to an Aptina MT9V024 CMOS sensor (752×488 pixels) and to a Spartan 6 FPGA from Xilinx, we designed a new embedded smart vision system (named « SmartNeuroCam ») which allows to track in real time and with a low power
consumption (less than 1w), human persons in natural scenes.