Abstract | ||
---|---|---|
In this paper we present a robust approach to real world, real time action classification. It relies on a convolutional network based object detector to extract relevant shape and motion features and uses these features as input for an action classifier. Using a sequence of localization and classification information of various objects deemed relevant to an action, the model recognizes predefined actions in a reliable manner, and can localize these actions in camera footage in real time. Without loss of generalization, we study our approach within the context of a construction company that wants to prevent unauthorized excavation activities happening at their construction sites. We differentiate four excavation activities, two of which we detect on the basis of actions because the target pattern contains temporal features, and two of which we detect on the basis of object presence only. The system needs to operate in real time, on basic on-site hardware and under varying image conditions. |
Year | Venue | Field |
---|---|---|
2018 | IDA | Excavation,Pattern recognition,Convolutional neural network,Computer science,Feature engineering,Artificial intelligence,Deep learning,Classifier (linguistics),Detector |
DocType | Citations | PageRank |
Conference | 0 | 0.34 |
References | Authors | |
14 | 5 |
Name | Order | Citations | PageRank |
---|---|---|---|
Bas van Boven | 1 | 0 | 0.34 |
Peter Van Der Putten | 2 | 72 | 8.84 |
Anders Åström | 3 | 0 | 0.34 |
Hakim Khalafi | 4 | 0 | 0.34 |
Aske Plaat | 5 | 524 | 72.18 |