Scale variant vehicle object recognition by CNN module of multi-pooling-PCA process

Abstract

The moving vehicles present different scales in the image due to the perspective effect of different viewpoint distances. The premise of advanced driver assistance system (ADAS) system for safety surveillance and safe driving is early identification of vehicle targets in front of the ego vehicle. The recognition of the same vehicle at different scales requires feature learning with scale invariance. Unlike existing feature vector methods, the normalized PCA eigenvalues calculated from feature maps are used to extract scale-invariant features. This study proposed a convolutional neural network (CNN) structure embedded with the module of multi-pooling-PCA for scale variant object recognition. The validation of the proposed network structure is verified by scale variant vehicle image dataset. Compared with scale invariant network algorithms of Scale-invariant feature transform (SIFT) and FSAF as well as miscellaneous networks, the proposed network can achieve the best recognition accuracy tested by the vehicle scale variant dataset. To testify the practicality of this modified network, the testing of public dataset ImageNet is done and the comparable results proved its effectiveness in general purpose of applications.

Keywords

• object detection
• scale invariance
• spatial pyramid pooling
• multi-pooling
• convolutional neural network (cnn)