Teach-in

The most straightforward approach to teaching vehicle classes to a computer would be presenting sample vehicles to it by running them across the sensor setup, entering the correct class for each one and having the learning process take place vehicle by vehicle. A procedure like this, however, remains only convenient as long as there are no mistakes happening and nobody changes his mind on how to classify a vehicle that has already been entered. Un-teaching particular informations while leaving the rest unaffected is extremely complex and so one would rather decide to start from the beginning than attempting to apply corrections.

Since mistakes can always happen and belated changes should always remain possible the CLASSAX teach-in process is designed a little different.



The first step is simply collecting vehicle data. This is done by means of a common terminal program set up to log the vehicle data which are output by the MD-440 via RS-232 to a file. Since in the respective output mode the MD-440 echoes received characters any typing is automatically logged also, and so the second step – adding the appropriate vehicle classes – can be accomplished at the same time.

The result obtained by this is usually a set of log files containing more or less good vehicle data along with more or less correct class information. These data can now be edited, corrected, selected and combined as desired before pesenting them to the CLASSAX learning algorithm. Plus, they remain available so teach-in can be repeated with modifications as often as required to create a really good class data set.



SENSOR LINE offers a comfortable tool chain running under Win9x/2000/XP to accomplish this task. CLXInput is a tool for viewing, editing, adding and deleting vehicle data from log files without having to perform cumbersome cut-and-paste procedures.



Having composed a file of perfect teach-in data this is fed into the actual learning algorithm. This is implemented in the tool MakeCLXC which produces a C source code file. The class data are expressed as constants in this file and still allow modifications by hand.



The C class data file is then used to create an executable program for the MD-440 master board (the slave board software is unaffected by class data). This is done by means of the free MSPGCC compiler/linker and related tools. The necessary program calls – normally also a quite cumbersome piece of work – are all made automatically by the tool CLX-MkMaster.



Loading this executable into the MD-440 master board is the final step. There are several programming devices available which only need to be plugged into the JTAG connector located at the edge of the board to write the program into its flash memory. The master will now perform AVC according to the loaded class data.

Usually the class data must undergo this procedure several times with repeated refinement and completion of the teach-in data until the required classification accuracy is obtained. The main effort during this phase consists of testing the performance and computing the statistical error rate. Tests must always be performed with independent vehicle data in order to see whether the data used for teach-in are already representative for the particular classes.