What is the loop detection?
Detection loop (Loop Closure Detection), visual SLAM problem, often pose estimation is a recursive process, i.e., the position and orientation of a current frame calculated by the solutions pose, wherein the error will therefore such a frame passed on, it is what we call the accumulated error. The reason we use a recursive before the next frame pose, as the two frames close enough, certainly the establishment of binding two, but far away from the two might not be possible to establish such a constraint relationship. This position can be established to identify historical frame pose constraints, that is, loop detection.
Loop significance of detection?
With the back-end visual odometry and SLAM-optimized front of the system, it seems to have been relatively easy to use. However, it has decreased in the same time improve the accuracy of real-time, once reduced accuracy, will face the problem long accumulated error, especially as ORB-SLAM as only local map optimization program. How do we balance this conflict? When people with minds may be incomplete impression of the scene before seen by comparison to detect the loop, and it is also possible for SLAM current frame and the last key frame similarity comparison, such as the similarity exceeds a certain threshold on It may be considered to be detected loop.
1. The bag of words model
Bag of words model is to feature as a word, consistency by comparing two images of words, to judge whether two pictures of the same scene.
In order to be able to feature classified as a word, we need to train a dictionary. The so-called dictionary that contains the set of all possible words, in order to improve the versatility, requires the use of vast amounts of training data.
Dictionary of training is actually a cluster of process. Assuming that all pictures CCP extracted 10,000,000 feature, you can use the K-means method put them together into a 100,000 word. However, if you just use these words to match the 100,000 words of the efficiency is still too low, because the need to compare each feature 100,000 times to find their corresponding words. To improve efficiency, the dictionary is built on the training process of a k-th branch, tree of depth d, an upper layer nodes provide a broad classification, the lower layer nodes provide a fine classification, until the leaf nodes. Using this tree, the time complexity can be reduced to the log level, feature matching greatly accelerated.
2. Based on the detection loop of CNN
Can detect key frame depth features, it can determine by comparing the similarity of the image depth features whether or not a loop, there are more robust, this technology is still developing convolution neural network.
3. Since the detection loop network coding
Still groping. . . .