Executive Summary

Background

The purpose of this project was to develop a new methodology to characterize truck body types along California Freeways. With new information on truck activity by body types, results from this study are expected to improve heavy duty vehicle classification in the Emission Factors (EMFAC) model and the California Vehicle Activity Database (CalVAD), and provide critical data that is required for the analysis of freight movement that will benefit the California Statewide Freight Forecasting Model (CSFFM) and other freight- or truck-related studies.

In this project, inductive signature technology was used to develop and deploy an advanced vehicle classification system comprising two distinct classification models at sixteen selected locations for two types of facilities: existing inductive loop detector (ILD) and weigh-in-motion (WIM) sites. ILD sites currently provide only vehicle volume counts. And although WIM sites provide axle-based truck classification and axle weight data, they cannot provide information on truck configuration that can provide further insight on industry and freight activity. Through this study, it is now possible to obtain higher resolution truck data which can enable more accurate estimates of GHG and other truck emissions, allow for decision makers to make more informed decisions for pavement management across a wider range of locations, and provide insight into the spatial distribution of body types for freight forecasting applications. The project will enhance CalVAD by incorporating a higher level of detail of commercial vehicle body classes, thus expanding the estimation of HHDT activity by CalVAD. It will also be used to improve the EMFAC model by increasing the number of truck types in the model and will improve on the understanding of truck activity in the State of California.

The project was divided into three phases with the following accomplishments: Phase 1 developed proof-of-concept body classification models; Phase 2 enhanced the proof-of-concept models and created techniques for obtaining body classification predictions for historical WIM data; and Phase 3 deployed the developed classification models to selected WIM and ILD sites located in California’s Central Valley.

Objectives and Methods

The research carried out under this project represents a completely new method for obtaining high resolution truck data. The objectives of Phase 1 included the (1.1) development of body type classification models using inductive signature data, (1.2) development of body type classification models using inductive signature data fused with WIM data, and (1.3) investigation of hardware interface configurations between WIM controllers and inductive signature technologies. Phase 2 objectives included (2.1) data collection, (2.2) model enhancement, (2.3) development of a methodology to propagate weight data to ILD locations, (2.4) development of a method to generate body class estimates from historical WIM data, and (2.5) development of an optimal facility location model. Lastly, Phase 3 was comprised of (3.1) equipment and model deployment to selected ILD and WIM sites and (3.2) system shakedown efforts for deployed sites.

Phases 1 and 2 required extensive data collection for model development and enhancement. Data was collected at seven WIM sites and one ILD site across California. Collected data – comprising inductive loop signatures, WIM controller outputs, and still images – were pre-processed, loaded into a relational database, and processed using a specially developed software user interface designed to assist with associating each collected sample’s data (signature, WIM data, and photo) and classifying the record according to a preliminary body configuration classification scheme. The truck body configuration classification scheme was originally derived from the body classes defined in the 2002 Vehicle Inventory and Use Survey (VIUS), and was further refined to reflect the variety of body configurations observed in the collected data. The proof-of-concept model for ILD-based classification included data from a single ILD site while the proof-of-concept model for loop and WIM focused only on the two most common truck classes: the FHWA class 5 single-unit two-axle trucks and FHWA class 9 five-axle semi-tractor-trailers. Unlike the proof of concept models developed during Phase 1 which employed a Feed Forward Neural Network model architecture, the fully enhanced models developed in Phase 2 followed a multiple classifier systems approach with probabilistic model combination to produce accurate and detailed truck body class predictions for all vehicle types ranging from two axle pickup trucks to six or more axle semi-tractor trailers. For the propagation of weight data to ILD sites (Task 2.3), a Gaussian Mixture Model (GMM) approach was employed in conjunction with Global Positioning Data (GPS) of truck trajectories to model the gross vehicle weight distribution of trucks at ILD sites by body type. For the backcasting task (Task 2.4), historical estimates of body class volumes of five axle semi-trailers were estimated from WIM axle weight, spacing, and length data using a modified decision tree framework while historical estimates of gross vehicle weight distributions were estimated by body class using GMMs in combination with the modified decision tree approach. Lastly, a method to select the deployment locations of the body classification models utilized an optimization model that selects the optimal site locations that capture the most number of unique truck trajectories in the GPS data set.

Results

The enhanced body classification models produced for WIM sites are subdivided by FHWA axle class such that each axle class has a unique set of body classes and resulting correct classification rates (CCR). Overall, the nine axle stratified models produce a total of 63 body classes with individual models ranging from four to 16 body classes. The CCRs were in the approximate range of 75% to 96%. The ILD body classification model was stratified into three tiers with the first tier predicting the general vehicle configuration of the vehicle as a single unit or multi-unit truck, the second tier predicting the body configuration, and the final tier predicting the body class. In summary, the ILD model consisted of 47 body classes across four body configuration groups with CCR ranging from 72% to 94%. The backcasting approach was applied to five axle semi-tractor trailers to predict volumes by trailer body type across five trailer classes: vans, tanks, platforms, intermodal containers, and others. The model produces body class volumes estimates with low error in the range of approximately 5% absolute percent error in volume. Finally, for the propagation of gross vehicle weight (GVW) to ILD site locations, accuracy of the model was determined by applying a hold-one-out method in which the GVW was estimated for an individual WIM site and compared to the observed GVW distribution at that site. This was repeated for all WIM sites. Approximately 65% of the predicted GVW distributions were found to be statistically significant representations of the observed distributions at the 95% confidence level.

The deliverables and results related to hardware development were demonstrated as part of the Phase 3 deployments to the sixteen selected sites (4 WIM and 12 ILD site locations) in the California San Joaquin Valley Air Basin. The deployment locations are shown in Figure ES-1. The deployment was performed on a variety of facility types, including freeways, highways and arterials.

Deployed Sites
Figure ES-1 Overview of site deployments

The live deployment of the classification models for ILD and WIM sites was completed using two servers hosted at the University of California, Irvine Institute of Transportation Studies. Data from the deployed sites were live-streamed, processed and archived into a relational database during the deployment period. The prototype Truck Activity Monitoring System (TAMS) web-interface was developed to provide access to on-demand detailed hourly summary reports of the body classification models at the deployed sites (sample queries results from TAMS are shown in Figure ES-2).

Deployed Sites
Figure ES-2 Example of detailed hourly volumes of FHWA class 9 trucks by body configuration at the Galt WIM site along the SR-99 freeway

Conclusions

This project has demonstrated and applied inductive signature technology to detect truck body class at inductive loop detector (ILD) and weigh-in-motion (WIM) sites in California. Extensive data collection efforts across the state resulted in an exceptional inventory of truck body types that were used to develop two truck body classification models: a standalone inductive signature only model and an integrated WIM and inductive signature model, which were designed for implementation at ILD and WIM sites, respectively. The standalone inductive signature only model is capable of distinguishing over 40 truck configurations, while the model combining WIM and inductive signature data is capable of predicting up to 63 body classes, both with overall accuracies above 70%. In addition to the body class models developed around inductive signature technology, the extensive dataset was leveraged to develop a methodology to predict body class volumes from WIM data for sites not equipped with inductive signature technology which is useful for backcasting tasks related to the validation and calibration of the California Statewide Freight Forecasting Model (CSFFM). Furthermore, the procedure designed to estimate gross vehicle weight distributions at ILD sites showed promising results. The deployment of inductive signature technology and corresponding body classification models to 16 sites in the California San Joaquin Valley gives practitioners and researchers a valuable tool to assess detailed truck activity, freight movements and impacts. The results obtained at selected sites have been shown to corroborate strongly with existing freight facilities in the region.

With new information on truck activity by body types, results from this study are expected to improve heavy duty vehicle classification in the EMFAC model and the California Vehicle Activity Database (CalVAD), and provide critical data that is required for the analysis of freight movement that will benefit the California Statewide Freight Forecasting Model (CSFFM) and other freight- or truck-related studies.

As a follow-up effort, Caltrans has sponsored a $1M study to further enhance the classification models and expand the number of deployed sites to over 90 across the State of California. These future deployments will be located along major truck corridors within metropolitan areas, at regional cordon lines, and near state boundaries. This follow-up study will also further enhance the TAMS web interface through which users can examine and download individual body class predictions in addition to hourly summaries by location.