Autonomous Multi-sensor Vehicle Classification for Traffic Monitoring

Abstract

AbstractThe world will see a tremendous increase in the number of vehicles on the road in the near future. Future traffic monitoring systems will therefore play an important role in improving the throughput and safety of our roads. Current monitoring systems capture traffic data from a large sensory network. However, they require continuous human supervision or a significant amount of hand-labeled data for training and both are extremely expensive.As part of a joint research project, we have investigated the scientific and technological foundations for future autonomous traffic monitoring systems. Autonomy is achieved by a novel combination of three approaches: self-learning and scene adaptive vision-based detection and classification, multi-sensor data fusion, and implementation on distributed embedded platforms.In this paper we present our self-learning and co-training framework with the goal of significantly reducing the efforts required for manual training in data labeling and autonomously adapting the classifiers to changing scenarios. Our system consists of a robust visual online boosting classifier that allows for continuous learning. We also incorporate an audio sensor as an additional complementary source into the training process. We have implemented the framework on an embedded platform to support mobile and autonomous traffic monitoring. We have demonstrated this by detecting and classifying vehicles based on real-world traffic data captured on freeways.KeywordsUnlabeled DataVehicle DetectionUnlabeled SampleTraffic MonitoringElectronic Toll CollectionThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.