Deep Extreme Learning Machines with Auto Encoder for Speed Limit Signs Recognition

This work presents a Deep Extreme Learning Machine with Auto Encoder scheme for Speed Limit Signs Recognition in the field of Advanced Driving Assistance Systems, where traffic sign recognition from video imaging plays an important role specially to provide vehicles with automated speed limits enforcement. Current solutions adopted by car manufacturers do not provide robust enough recognition behaviors when the image quality, the lighting conditions or the clearance of the traffic sign are compromised. These conditions result in misinterpreting of the speed limits, showing wrong on-screen advices which might confuse the driver, causing dangerous situations. In this work, the full chain of operations is studied. The proposed scheme is trained and tested with the German Traffic Sign Recognition Benchmark (GTSRB) database, achieving recognition times as short as 0.62 ms per sample, reaching with this timing real-time operation, and an accuracy of up to 92% with a simpler structure than other techniques currently used, such as Convolutional Neural Networks (CNNs).