Recognition performance at the DSP platform has accelerated from 150 ms to 80 ms and 90 ms

15 Jun 2016

The LPR team from FF Group has completed another stage in the acceleration of the recognition work process at the DSP platform. In cameras using a Texas Instruments DSP, the processing time of the FullHD frame decreased to 80-90 ms. 

This will increase the product significance for the customer: decreasing the recognition frame processing time and resources using, increasing the number of processed road lines for each camera.

Mykyta Sytyi, Team Leader of the ANPR project FF Group:

«Given the fact that we accelerated from the current 150 ms to 80 ms and 90 ms at the DSP, that is exactly possible that processes at the ARM CPU also can be slightly speed up. This can became possible when all changes to the DSP will be implemented in the main project branch

The work was carried out using iterative methods. The focus was on bottlenecks, which takes the most amount of time, and were worked out for the recognition acceleration. Meanwhile the quality of the DSP processing of the frame has not decreased. Several algorithms (translated from real numbers to integers), memory problems, and much more have been optimized during the acceleration process.

The next step of the developers command action will became the advising of all developed task branches in one stream of the code to simplify its support and possible acceleration at the ARM platforms. In particular, at Ambarella S2, S3 and S5 processors.