Analog Neural Network Design for Real-Time Surface Detection with a Laser Rangefinder

This paper deals with the design of an analog neural network (NN) that has to be implemented in a new kind of optoelectronic measurement system. The aim is to achieve real-time recognition of different kinds of surfaces using a phase-shift rangefinder and a neural network, at a remote distance between the rangefinder and the target varying from 0.5 m up to 2 m and with a laser beam incidence angle with regards to the target varying between -pi/5 and pi/5. The NN architecture is a multi-layer perceptron (MLP) with two inputs, three processing neurons in the hidden layer and one output neuron. Both inputs receive analog signals directly issued from the rangefinder. The output neuron gives a voltage that indicates which surface has been detected. The ANN design presented in this paper is completely analog in order to avoid signal pre-processing before the network and to limit power consumption. The 3.3 V voltage supply, relative to AMS CMOS 0.35 mum chosen technology, allows also a low system power consumption but important optimization of each circuit design is required. Each cell transfer is achieved with differential currents. Sigmoid function bias current is set outside the chip in order to control the tangent hyperbolic cell linearity range. NN global consumption is equal to 15 mA, with 4.4 mA per neuron. The neuron size is about 276 mumx 148 mum.