The walkability of a neighborhood impacts public health and leads to economic and environmental benefits. The condition of sidewalks is a significant indicator of a walkable neighborhood as it supports and encourages pedestrian travel and physical activity. However, common sidewalk assessment practices are subjective, inefficient, and ineffective. Current alternate methods for objective and automated assessment of sidewalk surfaces do not consider pedestrians’ physiological responses. We developed a novel classification framework for the detection of irregular walking surfaces that uses a machine learning approach to analyze gait parameters extracted from a single wearable accelerometer. We also identified the most suitable location for sensor placement. Experiments were conducted on 12 subjects walking on good and irregular walking surfaces with sensors attached at three different locations: right ankle, lower back, and back of the head. The most suitable location for sensor placement was at the ankle. Among the five classifiers trained with gait features from the ankle sensor, Support Vector Machine (SVM) was found to be the most effective model since it was the most robust to subject differences. The model’s performance was improved with post-processing. This demonstrates that the SVM model trained with accelerometer-based gait features can be used as an objective tool for the assessment of sidewalk walking surface conditions.

A neighborhood’s walkability is associated with public health, economic and environmental benefits. The state of the walking surface on sidewalks is a key factor in assessing walkability, as it promotes pedestrian movement and exercise. Yet, conventional practices for assessing sidewalks are labor-intensive and rely on subject-matter experts, rendering them subjective, inefficient and ineffective. Wearable sensors can be utilized to address these limitations. This study proposes a novel classification method that employs a long short-term memory (LSTM) network to analyze gait data gathered from a single wearable accelerometer to automatically identify irregular walking surfaces. Three different input modalities—raw acceleration data, single-stride and multi-stride hand-crafted accelerometer-based gait features—were explored and their effects on the classification performance of the proposed method were compared and analyzed. To verify the effectiveness of the proposed approach, we compared the performance of the LSTM models to the traditional baseline support vector machine (SVM) machine learning method presented in our previous study. The results from the experiment demonstrated the effectiveness of the proposed framework, thereby validating its feasibility. Both LSTM networks trained with single-stride and multi-stride gait feature modalities outperformed the baseline SVM model. The LSTM network trained with multi-stride gait features achieved the highest average AUC of 83%. The classification performance of the LSTM model trained with single-stride gait features further improved to an AUC of 88% with post-processing, making it the most effective model. The proposed classification framework serves as an unbiased, user-oriented tool for conducting sidewalk surface condition assessments.