| Summary: | 碩士 === 國立臺北科技大學 === 機械工程系機電整合碩士班 === 107 === The clarity of the windshield is a key factor in driving safety. However, the general public who use the wiper need to adjust the speed of the wiper according to the intensity of the rain, and very often are distracted. Nowadays, many domestic and foreign scholars have proposed optical raindrop sensing and capacitive raindrop sensing to detect the amount of rainfall, thus enabling the automatic operation of the wiper. However, the way to control the wiper may encounter problems such as the difference between the sensor and the surrounding environment, and it may not be close to the actual situation. The author believes that the timing of operation of the wiper is more related to the driver's visual perception of the raindrops on the windshield. Therefore, this paper attempts to directly judge whether the wiper should be activated by the image of the windshield and to use the convolutional neural network to perform image classification to achieve the effect of wise judgment. The author collects various driving video records of the streetscape through the windshield as a dataset, which is then manually divided into daytime and night and prepared as the required training set for the Convolution Neural Networks that command on the timing of wiper activation. The training accuracy of the integrated model of day and night is 99.75%. For verification of the models, we used untrained images to test. The precision of the daytime test was 93.7% and the recall rate was 79.8%. The precision for the nighttime test was 79.5% and the recall rate was 84.%. The calculation time required for each picture prediction process varies from 0.08 seconds to 0.13 seconds, which is much less than the shortest cycle time (about 0.8 seconds) that the wiper operates at the highest frequency.
|
|---|
