high performance computing on graphics processing units: hgpu.org

From the desire to update the maximum road speed data for navigation devices, a speed sign recognition and detection system is proposed. This system should prevent accidental speeding at roads where the map data is incorrect for example due to construction work. Multiple examples of road sign classification systems already exist but none uses a fully trainable solution. This feature enables the "vendor" to easily add new speed signs by training with a set of examples instead of designing a new system. To meet the above requirements a fully trainable Convolutional Neural Network (CNN) is used for the detection and recognition of speed signs. The system is trained with a labelled set of examples of speed sign images. Training of the total classification system is done off-line with the of error back-propagation algorithm. A trained system is used to collect new training data from road scene images to learn from previous errors, this technique is known as boosting. After the boosting step 0.19% of the images in our online available test set are misclassified. For the detection application the search window of the trained CNN is scaled to a 1280×720 HD image size to detect speed signs at multiple scales and positions in front of a vehicle. Because of the massive amount of parallelism in the computations of a CNN the algorithm is mapped to a Graphics Processing Unit (GPU). The GPU implementation demonstrates the abilities of the recognition system on a low cost consumer platform with a real-time frame rate of 35 fps.