| description abstract | The paper introduces a generic deep learning–based method using a sliding window computing algorithm based on long short-term memory (LSTM) networks for the classification of potential anomalies (e.g., cracks, potholes, bumps, and uneven surfaces) in support of the development of an instrumented bike. The instrumented bike provides a real-time platform to sense, store, transmit, and analyze cycling information through a sensor logger, smartphone, and the proposed LSTM-based sliding window computing algorithm. The paper is to address concerns with respect to existing factors such as weight of cyclists, speeds, types of bikes, and threshold setting that have an impact on the accuracy of identification of potential anomalies during instrumented cycling activities. The LSTM-based sliding window computing algorithm is designed in a way that it analyzes and localizes anomalies without any human-controlled supervision (threshold setting) while achieving human-level perception. Two bike routes were selected to validate the effectiveness of the sliding window computing algorithm in the identification of anomalies involving four cyclists. Based on the computing results from the two field tests, the numbers of distressed pavement areas from the four cyclists were 53, 51, 46, and 48, respectively. The follow-up p-value of ANOVA test result is 0.98, indicating the difference in detected anomalies among the four cyclists is not significant. Therefore, the paper concludes that the LSTM-based sliding window computing algorithm has the ability to effectively detect anomalies of cycling trails and it also provides an effective and efficient technique to replace the human-made threshold setting in support of the development of instrumented bikes and promote cycling as a daily mode of transportation. | |
