Real-time pothole detection on TMS320C6678 DSP

In this paper, a previously proposed pothole detection algorithm is improved in order to increase its performance in terms of speed and detection rate and to evaluate the performance. The algorithm was implemented for the TMS320C6678 SoC Digital Signal Processor (DSP). Potholes can be detected by comparing the disparity values of the scene to that of the disparity of the road surface if known. In a practical situation, the disparity of the road surface is not known but can be estimated. In this paper, the disparity of the road surface is efficiently estimated by modeling the road surface. Connected Component Labeling for size and severity estimation of individual potholes is also used. The main improvement of the existing algorithm follows the introduction of sampling when estimating model parameters. The sampling algorithm uses prior knowledge of the points' distribution and thus reduces the randomness when selecting data points. To achieve a high frame rate, the system was not only improved at the algorithm level but also taking advantage of the architecture of the SoC, using the various compiler optimization switches and selecting the appropriate memory arrangement. Finally, to take advantage of the multicore SoC, OpenMP APIs for code parallelization was used for the implementation. This has shown an improvement of around 3.4 times when using 5 cores. It is also shown that using more than 5 cores can result in performance degradation. Experimental results show that the system achieves real-time high performance of 145fps on images with 1050×600 resolution with 100% recall and 98% accuracy for pothole detection.