Automotive Interior Sensing - Anomaly Detection

Abstract

With the appearance of SAVs (Shared Autonomous Vehicles) there will no longer be a driver responsible for maintaining the car interior and well-being of passengers. To counter this, it is imperative to have a system that is able to detect any abnormal behaviours, e.g., violence between passengers, and trigger the appropriate response. Furthermore, the type of anomalous activities can be so diverse, that having a dataset that incorporates most use cases is unattainable, making traditional classification algorithms not ideal for this kind of application. In this sense, anomaly detection algorithms are a good approach in order to build a discriminative model. Taking this into account, this work focuses on the use of deep learning techniques, more precisely Spatiotemporal auto-encoder based frameworks, which are trained on human behavior video sequences and tested on use cases with normal and abnormal human interactions from Bosch's internal datasets. Initially, the model was trained on a single non-violent action category. Final iterations considered all of the identified non-violent actions as normal data. The network architecture presents a group of convolutional layers which encode and decode spatial data; and a temporal encoder/decoder structure, implemented as a convolutional Long Short Term Memory network, responsible for learning motion information. The network defines how to properly reconstruct the 'normal' frame sequences and during testing, each sequence is classified as normal or abnormal based on its reconstruction error. Based on these values, regularity scores are inferred showing the predicted regularity of each frame. The resulting framework is a viable addition to traditional action recognition algorithms since it can work as a tool for detecting unknown actions, strange/abnormal behaviours and aid in understanding the meaning of such human interactions.