Improving Indoor Pedestrian Detection and Tracking in Crowded Environments: Deep Learning Based Multimodal Approaches

Abstract

With the increasing demand for location-based indoor services and the rapid advancements in related technologies, indoor positioning has gained significant traction. This research dives into several leading indoor localisation technologies, introducing enhancements through different technologies. GauPro, a beacon probabilistic fingerprint algorithm, leverages the Gaussian kernel model and an LSTM model to refine the precision of conventional algorithms. Our system offers positioning services for standard smartphone users without requiring hardware modifications on the receiving devices. Computer vision-based indoor positioning grapples with challenges in object detection and tracking data association. Addressing this, we introduce GATracker, a two-stage pedestrian tracker. This model employs bipartite matching for tracklet proposals and a trainable graph attention network to pinpoint split points, enhancing tracking accuracy. To increase detection performance in densely populated scenes, we propose an anchor-free joint head and body detector JointTrack. This model dynamically learns the head-body ratio, eliminating the need for statistical data during training. Evaluations on datasets such as MOT20, Crowdhuman, and HT21 confirm its superior performance, especially for small and medium-sized pedestrians. The association of visual detection with users in crowded settings remains a challenge. Our solution PediFuse integrates real-time video analytics with wireless signal-enabled devices using a deep reinforcement learning scheme. The agent extracts motion patterns from both visual tracking and wireless signals, merging these features to identify the best match. In summary, we have taken a multifaceted approach to advance the state-of-the-art in indoor positioning systems. Each method we propose represents a significant step forward that addresses the distinct challenges faced in the field of indoor localisation.