Event-Cameras and Spiking Neural Networks for Tracking and Classification in Traffic Monitoring

Sammanfattning

Computer vision in real-world applications is an ever-growing field of great future and current potential. The main method for employing computer vision in these applications has been by using conventional frame-based cameras coupled with an Artificial Neural Network (ANN). A new promising technology, which has been growing in popularity, that could change this is Spiking Neural Networks (SNNs), together with event cameras. These technologies work in an event-driven fashion and promise real-time, energy-efficient data processing. This thesis will investigate whether an SNN could be trained to accurately identify and track multiple different types of traffic participants in a traffic monitoring situation. To achieve this, traffic image data captured by a conventional camera and an event camera, together with basic theory, were obtained from a previous thesis project, which focused on optimizing an SNN capable of tracking cars in traffic. The key result of this project is an SNN model which showed signs of being able to detect and track pedestrians, cars, buses, and trucks simultaneously, in a traffic monitoring scenario. The SNN model was trained with a supervised approach and evaluated by its Mean Squared Error loss value and by a visual inspection of the models output. Based on this result and the presented method for the result, further research can be made that either expands upon the presented SNN in terms of increasing the amount of tracked classes or further improving the models’ accuracy. The developed method that is presented may also be used to create a multi-object tracking SNN for use in another application.