Day 4

Detailed paper information

Back to list

Paper title Convolutional Neural Network and LSTM Applied to Abnormal Behaviour Detection from Highway Footage
Authors
  1. Rafael Marinho Instituto Nacional de Pesquisas Espaciais Speaker
  2. Elcio Shiguemori Instituto de Estudos Avançados
  3. Rafael Santos Instituto Nacional de Pesquisas Espaciais
Form of presentation Poster
Topics
  • C4. HAPs/UAVs
    • C4.01 Innovative UAV applications
Abstract text Relying on computer vision techniques and resources, many smart activities have been possible in order to make the world safer and optimized on resource management, especially considering time and attention as manageable resources. Once the modern world is very abundant in cameras, including especialy the ones on security cameras and military-grade Unmanned Aerial Vehicles or even affordable UAV in which are becoming more common on society. Thus, automated solutions based on computer vision techniques to detect, monitor or even prevent relevant events such as robbery, car crashes and traffic jams can be accomplished and implemented for the sake of both logistical and surveillance improvements, between other contexts and one way to do so is by identifying abnormal behaviours performed by the vehicles on observed roads. In this paper is presented an approach for vehicles’ abnormal behaviours detection from highway in which the vectorial data of the vehicles’ displacement are extracted from images captured by a stationary quadcopter UAV and surveillance cameras. Two deep neural networks used in this paper. A deep convolutional neural network was employed to object detection and tracking. Then, a long-short term memory neural network is used to behaviour classification. The deep convolutional neural network is a YOLOv4 trained with images extracted from highway footage, and the vehicles' vectorial data is extracted from their tracking on footages to train the long-short term memory neural networks. The training of the behaviour discriminator, in order to classify the behaviours as normal or abnormal, takes account the fact that most vehicles on the streets performs normal behaviours. The abnormal class is given by being an outlier on the general behaviours' profile. The results show that the classifications of the given vehicles' behaviours have been consistent and the same principles may be applied on other trackables objects and scenarios as well.