| Summary: | 碩士 === 國立臺北科技大學 === 電子工程系研究所 === 105 === With technology getting more and more developed, the Artificial intelligence and object detection skills are also growing very fast. Front-vehicle detection and lane departure products are getting more popular in the market, for example: alarm will warn driver when the distance is too close to front-vehicle or depart lane too much. Alarm’s reaction speed and correct rate will affect driver’s operation. Presume driving speed at 100/km, and alarm react only 1 second late will cause vehicle move 27 meters. Every second for high speed driving is very important. Therefore, this kind of product’s hardware and algorithm will directly affect driver’s safety.
Usually, vehicle-detection is a feature to catch car, for example: beneath shadow of vehicle, horizontal feature, vertical feature, lights, license plate, color, shape, features use more algorithms correct rate will be higher, but calculation time will be longer. Two most used features, like beneath shadow and horizontal feature are for vehicle-detection. But road traffic marking and words also have horizontal feature phenomenon, and that will cause vehicle-detection errors.
I’ve improved the traditional algorithm to against vehicle-detection errors when detecting road traffic marking. By using vehicle horizontal and beneath shadow features to do images overlapping, then take overlapping part with lane departure algorithm’s lane detection to detect road’s feature image after overlapping. By using feature’s shape, lane and car width relation to know is vehicle or not. Methods in this thesis could also decrease road traffic marking errors, and improve correction rate of front vehicle-detection based on sunny day.
|
|---|
