A Probabilistic Approach to Improve the Accuracy of Axle-Based Automatic Vehicle Classifiers

This paper details a simple and novel approach to solve the assignment problem of finding optimum thresholds for axle-based vehicle classifiers. A case study utilizing Oklahoma's axle-based classification stations was conducted in an effort to build, analyze, and verify the proposed solution. An extensive axle-space database with over 20 000 vehicle samples covering all 13 classes of the Federal Highway Administration's Scheme F was constructed. Histograms and Gaussian distribution fitting were individually done per class, per axle spacing, and from data gathered. Optimal class boundary thresholds were computed using the derived distributions, and a new classification algorithm was constructed. Results of field testing concluded that the newly proposed algorithm outperformed the existing one currently installed statewide and used by the other states as well. A significant false detection classification error reduction was achieved at a rate of 43% for class 8, 21% for class 5, 5% for class 6, and a combined reduction of 26% for classes 2 and 3. In addition, a misdetection error reduction of 21% for class 6, 13% for class 5, and a combined reduction of 37% for classes 2 and 3 was noted. A consolidated system error reduction relative to vehicle type was 15% for multiunit trucks of classes 8 to 13, 4% for single-unit trucks of classes 5 to 7, and 57% for passenger vehicles of classes 1 to 4. The process of calibrating the classification scheme was found to be completely transferable, thus could effectively be used to optimize classification algorithms in other states.