Identificação automática de rampas de acessibilidade apoiada por visão computacional a partir de imagens panorâmicas street-level

Abstract

Despite the legal provision for accessibility in urban public spaces, most sidewalks in Brazil are not endorsed with accessibility curb ramps. For people with reduced mobility, especially wheelchair users, both the absence of accessibility ramps and the lack of knowledge about their presence or not in a given location, restricts independent urban mobility and often discourages the movement of these agents, causing exclusion in the process of urban spaces democratization. The lack of information related to the location of these infrastructures is mainly due to the reduced availability of databases that cover them. Despite initiatives to map them, many methodologies become technically or economically unfeasible, such as the use of satellite images and field surveys. At the same time, the use of street-level images combined with artificial intelligence techniques, such as computer vision and neural networks, have been widely used to collect data on urban infrastructure. Seeking an alternative to this issue, the objective of this work was to use street-level images to build a labelled image dataset and enable the identification of accessibility curb ramps on sidewalks through object detection. To this end, the work began by obtaining Google Street View panoramas in a strategic way, based on census information on the occurrence of ramps and population strata. The accessibility curb ramps identified in these images were manually labeled for the construction of that image database and used in the training and validation of a convolutional neural network of the YOLOv4 object detector. From then on, training was performed varying pre-processing techniques and training parameters, thus, it was found that the use of Tiling and the use of pre-trained weights resulted in an average validation accuracy of the order of 65%. The main limitation of the research occurred in this stage, due to computational memory limitations for processing. Tests indicated that the network detects objects with an average accuracy of 85%, identifying about 77% of the objects in the test set. The resulting convolutional weights allowed ramps detection with varied designs, however, the neural network had a lower performance in the evaluation of partially occluded slopes or in a poor state of conservation. During the experimental procedures, it was observed that the adoption of the 65% threshold for detection accuracy led to a better balance between the number of correct and incorrect detections, as well as to enable the identification of proportionally small objects. The labeled image dataset elaborated in this work is a relevant contribution, considering that there is currently no similar database for accessibility curb ramps that includes Brazilian cities. In addition, the detector trained to identify curb ramps on public sidewalks proved effective, enhancing future applications that may involve the mapping of these infrastructures, the inclusion of other classes and the development of more elaborate applications.