Embedded Neural Network for Distance Recognition Using Distance Sensor

Determining the position of an object and avoiding the obstacles are crucial for robotic systems. The infrared and ultrasonic sensors are the main devices used for this purpose. Infrared (IR) Sharp GP2Y0A21YK0F with the range of 10-80cm was used in this work, which response in the analog voltage form. The relation between the distance measured and the output of the sensor is nonlinear. Hence, the output of the sensor needs to be converted to distance by one of the conversion methods. However, all conversion methods are imprecise hence they lead to occur large errors between the actual and measured distances. Artificial Neural Networks (ANN) were used to increase the accuracy of the conversion. Multilayer feed-forward (MLF) with different learning algorithms have been then implemented and tested. Instead of confining the ANN to a computer based application, this work extended the implementation of the neural network distance recognition in a general purpose microcontroller. The result was a low cost embedded system with a high accuracy in distance recognition. The experimental results demonstrated that the average percentage of ANN accuracy was 99.51% compared to 95.2% by a conventional method and the improvement of using neural network was very clear.