Comparison of Machine Learning Algorithms to Determine Traffic Congestion from Camera Images

For a country like United States where drivers spend $1200 per year on traffic jams, congestion detection is a subject of prime importance. In the present study images are obtained from CCTV cameras installed at different locations in the state of Iowa and five state-of-the-art shallow algorithms are explored to identify congestion. Features are extracted from the images using the OpenCV packages (Ski-Thomasi, ORB, and findcontours) and structured edge toolbox. The images are then segregated into test and training sets followed by combination of features using the training set. The two best performing ensembles are found to be OS (ORB and Ski-Thomasi) and OSS (ORB, Ski-Thomasi and Structured Edge Toolbox). The algorithms (k-NN, random forest and SVM) with variable parameters are applied on the training set to tune up the parameters. After that all the five algorithms - Naïve Bayes, k-NN, decision tree, random forest and SVM are applied on the test set. The results show that SVM has the highest f1-score of 86.73%. A sensitivity analysis is computed using receiver operating characteristic curve under time of day and camera conditions. It is seen that presence of glare or other inhibitions in the camera during the daytime leads to some misclassification of the traffic state. The results are within 5% of that obtained by deep learning algorithms. To conclude, the paper provides an option for a consumer to invest 60$ on a shallow algorithm and achieve an accuracy of 86.73% or to take up expensive deep model to improve the accuracy.

• Supplemental Notes:
  • This paper was sponsored by TRB committee ABJ70 Standing Committee on Artificial Intelligence and Advanced Computing Applications.
• Authors:
  • Poddar, Subhadipto
  • Ozcan, Koray
  • Chakraborty, Pranamesh
  • Ahsani, Vesal
  • Sharma, Anuj
  • Sarkar, Soumik
• Conference:
  • Transportation Research Board 97th Annual Meeting
  • Location: Washington DC, United States
  • Date: 2018-1-7 to 2018-1-11
• Date: 2018

Language

• English

Media Info

• Media Type: Digital/other
• Features: Figures; Photos; References; Tables;
• Pagination: 14p

Subject/Index Terms

• TRT Terms: Algorithms; Cameras; Detection and identification systems; Image processing; Machine learning; Sensitivity analysis; Traffic congestion
• Geographic Terms: Iowa
• Subject Areas: Data and Information Technology; Highways;

Filing Info

• Accession Number: 01658940
• Record Type: Publication
• Report/Paper Numbers: 18-04695
• Files: TRIS, TRB, ATRI
• Created Date: Feb 5 2018 11:24AM