The advancement in portable and lightweight sensor technology made the inertial data from daily activities widely accessible, contributing to the interest of tracking these movements or detecting characteristic patterns in these movements. This study focuses on detecting the crosswalk-crossing behavior of visually impaired individuals, which are observed to differ from those of normal sighted people. We hypothesize that a tactile search of the crosswalk endpoint using a cane, a hesitation at the endpoints, or a particular steady walking while crossing are reflected as a unique series of readings in the inertial data. Crosswalk-crossings generate rare and relatively long-duration patterns that regular walking actions, making it challenging to determine which features of the sensory data to focus on. We propose a specially tuned learning framework for detecting crosswalk-crossings using Deep Convolutional Neural Networks, which are known to be effective at data analysis involving temporal hierarchies. Considering also the use cases of the detected crosswalk-crossings, such as correction in the pedestrian positioning systems, we emphasize a precision-focused training with weighted F-scores and a related loss function that minimize false detections. Our findings show that our approach successfully detects real-world crosswalk-crossings with high precision and acceptable recall values. We also show that the models are capable of performing similarly on the data from different devices they are not trained with.

In inertial-only navigation systems, long-term position drift is inevitable without periodic resetting or anchoring. While numerous position improvement techniques have been proposed, the majority of the correction approaches rely on external data sources other than inertial measurements, introducing additional complexity and installation requirements. Furthermore, when attempting to anchor positions using pedestrian map information, it is often observed that pedestrians, including those who are visually impaired, often deviate from predefined walking paths, further complicating accurate positioning. To address these challenges, we introduce inertial-based correction technique for Pedestrian Dead Reckoning (PDR), with the aim of keeping all the processing in the inertial domain. We couple a personalized PDR solution for visually impaired pedestrians with a crosswalk crossing information detected in the inertial data. The correction technique relies on generating parameterized trajectory proposals that match the temporal information from detected crosswalk crossings with known crosswalk locations on maps. We demonstrate that this method reduces the positioning error of the underlying PDR by 50% on real-world data collected from visually impaired individuals, while also meeting real-time processing constraints.

Achieving accurate acoustic-based indoor positioning without synchronized clocks presents a significant challenge, particularly in collaborative systems where multiple devices must estimate pairwise distances simultaneously. In this paper, we propose a novel synchronization-free collaborative acoustic positioning system using standard smartphone speakers and microphones. Our system leverages a convolutional neural network (CNN) trained on Mel-frequency cepstral coefficient (MFCC) features for accurate pairwise distance estimation, followed by a lateration algorithm for position calculation. We demonstrate the feasibility of our approach through real-world experiments, achieving sub-meter positioning accuracy in the best-performing configuration (RMSE = 0.50 m, 90th percentile = 0.80 m). Other configurations, tested under more challenging conditions (e.g., long distances and less optimal geometric placements), exhibit higher errors. Our results highlight that low-frequency tones (440 Hz and 1000 Hz) provide more robust performance, while higher frequencies excel in mid-range distances but at near and far distances are more susceptible to environmental factors decreasing their reliability. Overall, the results indicate that our approach offers a promising solution for practical, synchronization-free acoustic positioning in collaborative systems.

The popularity of lightweight portable sensors has made inertial data from everyday activities widely available, sparkling interest in tracking these activities. This study focuses on detecting the motion patterns of visually impaired people while crossing a crosswalk, which is observed to be different from sighted people. Visually impaired individuals exhibit different behaviors, such as using a cane to find a crosswalk terminal, resulting in unique patterns in inertial data. This work proposes a crosswalk detection method based on inertial signals captured from naturally walking visually impaired pedestrians. Crosswalk crossings are long-lasting actions and rare events in a walking trajectory, introducing questions on which features to use and which parts of the data to investigate. We handle this problem by efficiently tuning Deep Convolutional Neural Networks with their multi-resolution feature extraction capability. Considering also the detected crosswalks in the correction of dead-reckoning based positioning applications, we state that the false detections are intolerable. Thus, we also perform precision-driven training processes, in which the weighted F-score and related surrogate loss functions are employed. The results demonstrate that the proposed approaches detect crosswalks effectively and show potential in classifying other similar long-duration actions.

Visually impaired people face particular challenges when it comes to positioning and navigation. This paper presents an innovative assistive technology designed to improve the independent mobility of visually impaired people by integrating multiple navigation inputs into a cohesive real-time positioning system. At the heart of our approach is the optimized use of an asynchronous particle filter — a probabilistic, recursive algorithm tailored to pedestrian navigation using Pedestrian Dead-Reckoning (PDR) and Map Matching. Unlike traditional mesh graphs, we employ pathway graphs specifically designed for visually impaired people, incorporating landmarks as critical nodes to enhance accuracy and flexibility. This design adapts to real pedestrian movement and improves the practicality of the system on walkable paths by allowing deviations from theoretical paths. Our system integrates graph edge observations, GNSS data, and landmarks to achieve high localization accuracy and robustness. We have evaluated three advanced navigation models that show significant improvement in stride trajectory estimation accuracy. Model 1, using GNSS, reduced the mean and median Euclidean distance to ground-truth to 2.56 meters and ${2. 2 6}$ meters, respectively. Model 2, which incorporates simulated landmarks, achieved mean distances of 2.17 meters and a median of 1.97 meters, while Model 3, which uses a graph edge-based approach, achieved the best results at a mean of 1.7 meters and a median of 1.47 meters. These results underline the improved accuracy of our navigation systems for visually impaired users, especially in complex environments.

The term "sector" in professional business life is a vague concept since companies tend to identify themselves as operating in multiple sectors simultaneously. This ambiguity poses problems in recommending jobs to job seekers or finding suitable candidates for open positions. The latter holds significant importance when available candidates in a specific sector are also scarce; hence, finding candidates from similar sectors becomes crucial. This work focuses on discovering possible sector similarities through relational analysis. We employ several algorithms from the frequent pattern mining and collaborative filtering domains, namely negFIN, Alternating Least Squares, Bilateral Variational Autoencoder, and Collaborative Filtering based on Pearson’s Correlation, Kendall and Spearman’s Rank Correlation coefficients. The algorithms are compared on a real-world dataset supplied by a major recruitment company, Kariyer.net, from Turkey. The insights and methods gained through this work are expected to increase the efficiency and accuracy of various methods, such as recommending jobs or finding suitable candidates for open positions.

Knowing the salary range of a position is beneficial for both job seekers and employers. This work examines the performance of different machine learning methods on salary estimation using industrial variables. The methods are applied to a dataset obtained from Turkey’s largest employment platform Kariyer.net. We perform various exploratory analyzes of the data, then use feature engineering techniques for improving the quality of the training data. The effect of the heavy-tailed distribution of salaries is mitigated with various response variable transformations. A timeliness standardization is performed using inflation rates as data from different time periods. Analyses and experiments show that standardization does not have a significant effect on the performance of the model. On the contrary, response variable transformation seems to have a significant effect. As for the models, we conclude that the XGBoost and the artificial neural networks achieve the highest success.

This work investigates the performance improvement of an indoor positioning and tracking system through a real-time preprocessing of the measured received signal strength indicator (RSSI) data. The system itself is based on a hidden Markov model constructed upon a priorly estimated radio frequency map for the measurement density, and a Gaussian (diffusion) distribution for the transition density. The positions are estimated as the latent variables via particle filter algorithms that are fed with the filtered RSSI data as observations. We first compare the three nonlinear time window filtering techniques, mean, median and maximal filters on the streaming RSSI data captured by the distributed Bluetooth low energy (BLE) sensors. Seeing the performance boost of the maximal filter strategy with a standard particle filter implementation, we further investigate various model parameters in two particle filter applications: static and adaptive particle filters. The maximal filter preprocessing technique is shown to increase the positioning performance by more than 20% for real-time applications. The performance boost has still space to perform 40% better with better approximation preferences compared to raw RSSI readings.

We construct a practical and real-time probabilistic framework for fine target tracking. The practicality comes from the application of the forward algorithm and the small parameter set used to build the hidden Markov model (HMM). A Bluetooth Low-Energy (BLE) beacon navigating in the environment publishes BLE packets which are captured by the stationary sensors. Fingerprints are formed by collecting received signal strength indicators (RSSI) of these packets, which are then processed into a high resolution emission matrix using a histogram combination technique. We convert the map of the area into a grid structure, the resolution of which is controlled by the grid cell size. The transition matrices are built by Gaussian blur masks parametrized by the size and diffusion extent. As the transition matrix is highly sparse, we make the exact inference tractable by adopting a sparse matrix representation and by intelligently controlling the mask size, diffusion factor and grid cell size. Filtering can then be directly performed by the forward algorithm given a series of real RSSI measurements along real trajectories. We measure the performance of the system by comparing the most likely positions at each step with the ground truth positions. We achieve promising results and evaluate the approach also by the runtime and memory usage.