Research on small target detection based on improved Faster R-CNN depth network method

Abstract

In order to solve the problem of detecting the cone barrel targets on both sides of the unmanned formula racing car, the traditional method is to extract features based on artificial design and select appropriate classifiers for detection and identification. This method is not ideal for detecting small targets of cone buckets. It is difficult to meet the requirements in terms of accuracy and speed of detection, and there are very large limitations. A new method based on improved Faster R-CNN for cone bucket target detection is proposed. This method can effectively avoid the problem of relying on artificial design features and classifier selection, and the speed and accuracy of detection are significantly improved. Method The method is based on a deep convolutional neural network. First, we have to determine the target object to be detected: the cones with the same shape and size on both sides of the track but different colors can be divided into red, yellow and blue cones according to the color. Next, the data set of the cone bucket target is created, and the corresponding configuration file is generated according to the format of the VOC 2007 data set. The sample data in the training set is sent to the network to obtain a convolution feature map. In order to fully extract the characteristics of the sample data, the method uses a deep residual network as the feature extraction network. After obtaining the convolution feature map, it is sent to the Region Proposal Network (RPN) for generating anchor boxes. Distinguish whether the target object belongs to the foreground or the background and perform a rough position correction on the region proposals. In order to enhance the detection effect on the small target cone, the finer division of the anchor boxes in the RPN. The traditional image processing method is used to realize the Region of Interest pooling (ROI pooling), and the output region proposals is fixed to a unified scale. The region proposals are extracted by optimizing the non-maximum suppression (NMS) algorithm to reduce the missed detection rate of the adjacent cone bucket target. Finally, the Softmax classifier is used to judge the category to which the target object belongs and to use the bounding box regression to achieve accurate position correction. Result The network model of target detection was obtained by using four-step training on the self-made cone bucket dataset. We compare the original Faster R-CNN model with the improved Faster R-CNN model to test the detection effect under different weather and environment. The improved model improves the detection of mutual occlusion between small targets and adjacent targets. The average accuracy on the test set is as high as 88.1%, and the detection frame rate reaches 10FPS. The improved method for cone bucket target detection proposed in this paper avoids the problem that traditional target detection methods need to manually select features and time-consuming. There is an obvious improvement in the accuracy and speed of detection, and it has strong robustness. It provides a new method for detecting the target of cone bucket of formula racing car.