Image-only place recognition based on regional aggregating ConvNet features for underground parking lots

Place recognition searches the closest map node to the query node, which is an important task for vehicle localization. Traditional visual place recognition methods for underground parking lots require the deployment of additional location signals, such as WiFi, Bluetooth. This paper utilizes only front-view images to realize place recognition. First, we employ a random coefficient to reduce the dimensionality of the ConvNet features to obtain the CCFs (Condense ConvNet Features). Second, we average the CCFs of a regional zone to obtain the RACF (Regional Aggregating ConvNet Feature). Compared with WiFi, Bluetooth, RACF is extracted from the front-view image and has a superior ability to represent regional zones. Third, we propose a multiscale place recognition method that adopts a coarse-to-fine strategy that greatly reduces time consumption and accelerates precision. Finally, we evaluate the proposed method on the data collected in the underground parking lot of Hubei University of Technology. The experimental results illustrate that the proposed method has high precision with a fast speed.