Detection of Anomalies in Pedestrian Walking Using OpenPose and Bidirectional LSTM

OpenPose is employed for accurate human pose estimation, while bidirectional LSTM is used to model temporal dependencies in pedestrian movement. Despite its potential, this research faces several challenges. Firstly, the accurate detection of subtle anomalies in pedestrian gait requires robust feature extraction and representation. Secondly, handling occlusions, varying walking speeds, and complex environmental conditions can impact pose estimation accuracy and anomaly detection. Additionally, the scarcity of annotated anomaly data poses a challenge for model training and generalization. Temporal Convolutional Networks (TCN), Attention mechanisms, LSTM are recently used techniques. This work deals with the processing of camera recordings of pedestrians walking and the subsequent detection of abnormal events such as a person falling while walking. To achieve this functionality, the bidirectional LSTM is used that detect people in the image and extract the coordinates of their skeleton. The study will subsequently use the data obtained from the human skeleton to detect anomalies using Openpose.