Road surface traffic sign detection with hybrid region proposal and fast R-CNN

Fast and robust road sign detection in driver assistance systems

“…However, the performance of its method degrades when there exists low image contrast or various illumination conditions. Note that, the method in [37], [21] and [33] performs comparably with the method in [8], with a slight improvement. Because the dataset contains more mandatory road signs, the performance of the method in [23] drops slightly.…”

“…In order to demonstrate the superior performance of our proposed algorithm, we compare the performance of our algorithm with those of the state-of-the-art approaches on three benchmarking datasets and a newly collected dataset. These approaches include the SVM approach proposed in [26], the approach integrating object detection with 3D tracking in [37], the Fourier descriptors in [21], the ROI extraction and histogram features-based method in [23], the image segmentation and shape analysis method in [20], Mean shift and log-polar transform [8], fast R-CNN method [33] and the RST approach [24]. Detection rate (Hits), false alarm rate (FA) and average processing time are compared.…”

“…[21] 95.1% 1.27% 2s Liang [23] 95.8% 1.03% 800mš Segvi [36] 95% 1.28% -Timofte [37] 94.7% 1.34% 286ms Ellahyani [8] 94.9% 1.3% 389ms Qian [33] 96% 1.27% 950ms [37] 95.6% 1.61% 3s Larsson [21] 95.4% 1.26% -Liang [23] 94.9% 1.28% 986mš Segvi [36] 92.9% 2.04% 451ms Ellahyani [8] 95.1% 1.74% 419ms Qian [33] 95.8% 1.39% 1.2s Table 2 shows the comparative results obtained on this dataset. As it shows, the results obtained using the method in [23] and [33] are very close to that of our algorithm, but the method in [23] and [33] run at a frame rate of 0.5 frames and 0.95 frames per second respectively, whereas our system runs at a frame rate of approximately 5 frames per second on this dataset. The processing speed in [37] is very close to our method, about 0.2 ∼ 0.4 second per image, but its detection rate is not as comparable as ours.…”

“…Ours 97.1% 0.75% 318ms Larsson [21] 96.8% 0.94% 2.4s Khan [20] 92.3% 1.95% 32s Liang [23] 96.5% 1.02% 988ms Loy [24] 94.4% 0.98% 496ms Bascn [26] 93.8% 1.24% 608ms Timofte [37] 92.6% 1.71% 351ms Ellahyani [8] 96.1% 0.89% 408ms Qian [33] 96.7% 0.99%…”

“…However, it did not mention the use of other methods for classification and recognition processes. [33] proposed a road surface traffic sign detection system by applying convolutional neural networks (CNNs). Extensive experiments have been conducted using their field-captured dataset, demonstrating outstanding recall and precision rate.…”