Adaptive background subtraction technique with unique feature representation for vehicle counting
Vehicle detection is the first step towards a successful traffic monitoring system. Although there were many studies for vehicle detection, only a few methods dealt with a complex situation especially in traffic jams. In addition, evaluation under different weather conditions (rainy, foggy and snowy...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2022
|
| Subjects: | |
| Online Access: | http://eprints.utm.my/id/eprint/101678/1/MohammedAtefAbdElMoniemPSKE2022.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Vehicle detection is the first step towards a successful traffic monitoring system. Although there were many studies for vehicle detection, only a few methods dealt with a complex situation especially in traffic jams. In addition, evaluation under different weather conditions (rainy, foggy and snowy) is so important for some countries but unfortunately it is rarely performed. Presently, vehicle detection is mainly performed using background subtraction method, yet it still faces many challenges. In this thesis, an adaptive background model based on the approximate median filter (AMF) is developed. To demonstrate its potential, the proposed method is further combined with two proposed feature representation techniques to be employed in either global or local vehicle detection strategy. In the global approach, an adaptive triangle-based threshold method is applied following the proposed adaptive background method. As a consequence, a better segmented foreground can be differentiated from the background regardless of the different weather conditions (i.e., rain, fog and snowfall). Comparisons with the adaptive local threshold (ALT) and the three frame differencing methods show that the proposed method achieves the average recall value of 85.94% and the average precision value of 79.53% with a negligible processing time difference. In the local approach, some predefined regions, instead of the whole image, will be used for the background subtraction operation. Subsequently, two feature representations, i.e. normalized object-area occupancy and normalized edge pixels are computed and formed into a feature vector, which is then fed into the k-means clustering technique. As illustrated in the results, the proposed method has shown an increment of at least 10% better in terms of the precision and 4.5% in terms of F1 score when compared to the existing methods. Once again, even with this significant improvement, the proposed method does not incur noticeable difference in the processing time. In conducting the experiments, different standard datasets have been used to show the performance of the proposed approach. In summary, the proposed method has shown better performances compared to three frame differencing and adaptive local threshold methods. |
|---|