Vision-Sensor Enabled Multilayer CNN Scheme and Impact Analysis of Learning Rate Parameter for Speed Bump Detection in Autonomous Vehicle System

Detection of speed bumps is a challenging issue in autonomous ve-hicle systems due to their appearance in dynamic environments. The unmarked speed bump is not an apparent entity even in the daytime, thus carries difficulty to precisely detect speed bumps for a vision-based driving assistance system (DAS). In some cases, DAS may misinterpret a marked speed bump and a zebra crossing pattern, which can further weaken recognition performance and may risk human safety. In this letter, a vision-based convolutional neural network model using a region proposal and a nonmaximum suppression speed bump detection network (SBDNet) is proposed for the accurate detection of speed bumps on the road. The impact of the learning rate has also been highlighted and analyzed during the experiment. The performance of the proposed SBDNet has been validated on the publicly available dataset “Speed Hump/Bump Dataset.” The overall performance of the proposed SBDNet has been achieved in terms of accuracy, sensitivity, and F-measure as 99.34%, 99.52%, and 99.04%, respectively, under the heterogeneous environment.