Research On Multi-scale Vehicle Detection

Object detection technology is a useful technology, but also a difficult and challenging technology. In many objects detection, vehicle detection is more difficult. It's mainly due to vehicles have a variety of models, such as trucks, cars, trucks, buses, etc. with different colors, but also with the different angle very different.A lot of object detection technology could be use to the vehicle detection. The Haar-Like features which be trained with adboost classification learning algorithm, show almost perfect reslut in face detection, and has some effect on vehicle too. But the variability of the vehicle make it not very good. The Gabor feature has a good effect in the vehicle detection, because it's not as sensitive to the gray as Haar-Like feature. The HOG feature, however, has good performance on both vehicle and pedestrian detection. Because it's main based on the margins in the photo, and it's robust. The large computation is a serious problem in the detection, and the cascade classifier could reduce a lot of computation.Vehicle detection has a common problem, the vehicle aspect ratio (referred to here as multi-scale) ofen appear quite different, such as the front face and rear face of the car is near square, and the side face of the car is near rectangular, it appears much longer for the bus. Inorder to detect different scales cars, it need to train classifiers with different aspect ratios. While intensive scanning the photo, multiple classifiers are need to classify. And then it need to merge the result windows of the different classifiers use a merging algorithm, so that it can achieve multi-scale vehicle detection.