Structural Detection of Visual Objects for Mobile Robots

Abstract

This paper presents StructDetect, a fast method for object detection. The target detection process consists of two stages: generation of a hypothesis (object proposals) (1) and verification of the hypothesis (2). Generation of the object proposals is carried out by means of a simple structural model on the basis of line segment combining. Line segment is detected by EdLines algorithm. Then a computer attributes the line segments and their pairs and creates "a connection table", which filters some combinations. Further, it creates a triple combination of the line segments filtered by "the connection table". Each combination has a handcraft descriptor based on the line segment attribute. This descriptor is used to learn kNN classifier and generate object proposals in the area of 3 line segments. These proposals define a set of candidate bounding boxes available to the detector. The second module is based on a convolutional neural network, which takes a fixed-length feature vector from each region. The convolution neural network computes once per image and features vector extracts with adaptively-sized pooling from the last convolution layer. Then the feature vectors are classified by the random forest algorithm. Accuracy of this approach is comparable with the accuracy of such modern detector methods as SPPNet and RCNN. StructDetect is 7 times faster than SPPNet and has a frame rate of 4fps on a CPU.