each window candidate will be passed through feature extraction module then the computeddescriptor will be passed into the corresponding binary classifier as shown Fig. 12. For detailsof each method, the readers are invited to read the original papers.In this work, we are interested to detect and recognize four classes of obstacles: {Pottedplant, Trash, Extinguisher, and Human}. For training and testing detection and recognitionmethods, we have built a dataset containing 2104 images. The resolution of images is 600x400pixels. Each object class has 526 images under daylight condition in a corridor of a build. Thisdataset is very challenge because objects are taken under different views point and distances.Some examples are presented in the Fig. 13. All images in the database are annotated manuallyand organized in the directory. We divide the database into 2 parts: 504 images for trainingand 1600 images for testing.To evaluate detection and recognition algorithms, there are many measures such as Recall,Precision, and Accuracy [8]. In our context, as we know the distribution of positive andnegative examples (the ratio between positive and negative is 1/4) so we propose to evaluateour system by Precision criterion, which is defined as follows:
đang được dịch, vui lòng đợi..
