A neural network based place recognition technique for a crowded indoor environment

Place recognition in a crowded and cluttered environment is a challenging task due to its dynamic characteristics such as moving obstacles, varying lighting conditions and occlusions. This work presents a robust place recognition technique that could be applied into a similar environment, by combining well known Bag of Words technique with a feedforward neural network. The feedforward neural network we use have three layers with a single hidden layer and it relies on rectifier and softmax activation functions. We employ cross entropy function to model the cost of our neural network and utilize Adam algorithm for minimizing this cost at the training phase. The output layer with softmax activation in the neural network, produces a vector of probabilities which represent the likelihood of test image being captured from a given region. These values are further improved by incorporating a transition matrix which is based on the building layout. We have evaluated our neural network based place recognition technique with data collected from a crowded indoor shopping mall and promising results have been observed by this approach. We also have analyzed the behavior of neural network for changes in hyper-parameters and presented the results.