Implementation of different image edge detection algorithms on a real embedded ADAS platform

Advanced Driver-Assistance Systems (ADASs) are becoming more and more popular in modern vehicles in the last years. A number of them are based on processing the images captured by in-vehicle cameras. Furthermore, in image processing-based ADAS algorithms one of the most important steps often is edge detection. Therefore, it is important to properly choose the edge detection method, to achieve high performance and low processing time. Due to limited hardware resources available when using real ADAS platforms, the trade-off is needed. In this paper, the implementation of four different edge detection operators (Sobel, Prewitt, Laplace and Canny) onto a real ADAS Alpha board is performed. The implementation is performed using the Vision Software Development Kit (SDK), which is a multi-processor software platform specifically optimized to work with Texas Instruments (TI) Systems-On-Chip (SoCs) that Alpha board consists of. To test the operators’ performance in a real ADAS operational environment, the Berkeley dataset is used. The output results of each edge detector are compared to available ground truth labeled images from the Berkeley dataset to check which detector achieves the highest performance in terms of edge detection accuracy. Furthermore, the operators are compared in terms of execution time and memory usage. It was shown that Canny operator requires the longest execution time and the highest amount of memory, but it also achieves the highest edge detection accuracy. It is also shown that the trade-off between detector accuracy and its requirements can be achieved in certain situations where it is acceptable.