Which is the Greenest Way Home? A Lightweight Eco-Route Recommendation Framework based on Personal Driving Habits Huihui Chen, Bin Guo, Zhiwen Yu Northwestern Polytechnical University Xi’an, China
[email protected], {guob, zhiwenyu}@nwpu.edu.cn
Alvin Chin, Jilei Tian
Chao Chen
BMW Technology Corporation Chicago, USA
[email protected],
[email protected]
Chongqing University Chongqing, China
[email protected]
Abstract—A vehicle’s fuel consumption is strongly related to both its loading and the driver’s driving behavior, such as aggressive/tender acceleration, improper/proper gear change or running/stopping the engine while waiting. This paper introduces EasyRoute, an economical route recommendation system for modern vehicles, implemented in smartphones, to improve fuel efficiency. EasyRoute senses the vehicle’s fuel consumption through the On-Board Diagnostics (OBD) adaptor and then models the driver’s personal fuel consumption according to OBD data from two aspects: (i) when the vehicle is moving and (ii) when the vehicle is idling or waiting. Based on the crowdsensing traffic information, EasyRoute can near-correctly predict total fuel consumptions of different routes and recommend drivers with the greenest route. We describe the EasyRoute framework and evaluate it by collecting OBD and GPS data from 559 taxis in Beijing. Comparing with some commonly used baselines with error metrics, the experimental results show that using a short 10minute dataset for training, the total fuel consumption estimated by EasyRoute has a relative error of at least 30% less than the baselines.
recommendation becomes another interesting research topic, as exploited in GreenGPS [11]. This paper focuses on eco-routing recommendation. We present EasyRoute, which utilizes OBD (On-Board Diagnostic, installed in the vehicle) readings and GPS trajectories of trips to predict the total fuel consumption of the new trip and then recommends the most economical route to the driver. In order to revise the route during the trip, EasyRoute trains the fuel consumption model (FCM) on driver’s smartphones in real time, therefore, we propose a light-weight method to predict the fuel consumption. In order to build EasyRoute we have the following challenges: On one hand, there are a large number of types, categories and makes of vehicles, and the same type of old vehicles might have different fuel efficiency. On the other hand, the average fuel consumption might change if driving contexts change, such as opening the air condition, carrying different numbers of passengers or different tons of goods, and so on. Therefore, we will not collect parameters of the vehicle, and the FCM trained by EasyRoute will not be directly related to static parameters of the vehicle.
Keywords—Personalized route recommendation; economical route; On-Board Diagnostic; smart phone; fuel consumption prediction.
The driver cannot provide historical fuel consumption information of all roads (e.g., the number of road segments in Beijing according to OpenStreetMapa (OSM) is more than 80,000), so the fuel consumption will not be related to roads but be related to traffic conditions of roads.
I. INTRODUCTION Much research has been done to find ways to reduce the fuel emission from vehicles by reducing the fuel consumption and they can be summarized from two sides: i) analyzing the car’s capability in order to save fuel as well as reduce the emission [14]; ii) proposing fuel consumption models that are related to different parameters of both drivers’ driving and vehicles [3, 57].
The total fuel consumption of a trip is predicted and revised in real time, in order to save the traffic the smartphone can undertake some pre-processing of data, so we will design a lightweight algorithm to save computation. To solve above challenges we only collect the vehicle’s selfdiagnostic data such as instantaneous speed, accumulated mileage, accumulated fuel consumption, and GPS data over OBD (On-Board Diagnostics). We then build the FCM on the smartphone using two parts: i) fuel consumption while moving; and ii) fuel consumption while waiting. According to the crowdsensing traffic information [12] and the FCM, we propose
In order to reduce the total fuel consumption of a trip studies can be broadly grouped into two directions: eco-driving and ecorouting. Eco-driving aims to recommend drivers to drive at proper speeds or exhibit good driving behaviors [7-9], while eco-routing is about providing the most economical route [10, 11]. Because the fuel consumption can be influenced by human driving behaviors and the vehicle's capabilities, personalized a
http://www.openstreetmap. org/
1
a framework to predict total fuel consumptions under different routes, and recommend the most economical route to the driver. Summarily, our work makes the following contributions:
(liter/km)). In this paper, the total fuel consumption calculation is based on both the time-based methodand the distance-based method. Another factor to compute the total fuel consumption is the traffic condition [14, 16]. GreenGPS [11] uses a set of static street parameters (e.g., the number of traffic lights and the number of stop signs), a set of dynamic street parameters (e.g., the average speed on the street or the average congestion level), and the vehicle parameters (e.g., weight and frontal area). Shang et al. [4] considered traffic volumes of different roads. Our work is different in that we use a crowdsensing dataset that consists of a sequence of speed-distance pairs or waiting timespans to describe the traffic information, from which they are used by our model to estimate the fuel consumption of different trips.
We create the personalized FCM for different drivers, suitable for mobile devices, and is constructed based on OBD observations (including the vehicle speed, the travel distance and the total fuel consumption) that implicitly reflects drivers’ behaviors and the fuel consumption ability of the vehicle. We propose a method to estimate total fuel consumptions of a trip via different routes using real-time crowdsensing traffic information and the personal FCM conducted on the smartphone, and then recommend the most economical route to the driver.
B. Methods to Save Fuel There are two methods for saving fuel: eco-driving [3, 5, 79, 15,] and eco-routing [6, 10, 11, 16, 17].
We collect a driving dataset of OBD and GPS data from 559 taxis in Beijing in order to validate the FCM and the total fuel consumption prediction methods.
Eco-driving aims to recommend drivers to drive at proper speeds or exhibit good driving behaviors, such as tenderly starting, less change of speed, early accelerator off, and proper gear change moment, when they are driving on different roads to achieve fuel savings. Hu et al. [8] introduced a crowdsourcing-based vehicular social networking system for green transportation, which used a rating mechanism about the fuel economy of drivers, provided a user-friendly mobile application on smartphones and promoted drivers’ awareness of their driving behaviors regarding fuel economy of specific trips.
The rest of the paper is organized as follows. We review related work in Section II, and Section III overviews the framework. We propose the FCM and detail the fuel consumption prediction methods in Section IV. We evaluate our approach and discuss the implications of the results in Section V and conclude the paper in Section VI. II. RELATED WORK Much research has examined various methods for estimating fuel consumption and saving fuel. We explain the related work below and point out how our work differs from them.
Eco-routing is about providing the most economical route. To our knowledge, the recent eco-routing work is GreenGPS [11], which used OBD data and found the most fuel-efficient route for drivers. Either directed by vehicle manufacturers or via OBD devices, GreenGPS is typically vehicle-specific and requires extra devices/efforts. Differences between EasyRoute and GreenGPS include: i) In order to train the model, GreenGPS collects instantaneous vehicle speed, mass air flow, command equivalence ratio, engine RPM, throttle position, latitude, longitude, altitude, bearing, and time, while EasyRoute only collects instantaneous vehicle speed, accumulated fuel consumption, latitude, longitude, and time; ii) GreenGPS opportunistically uploads (e.g., via free WiFi) OBD sensor data and location data, while EasyRoute uploads a series of preprocessed traffic information (around 1kb per 10 minutes) and the FCM (around 200 bytes) in real time; iii) EasyRoute trains the FCM and senses the traffic condition on smartphones using low-frequency (0.1 Hz) OBD readings and GPS readings and only the complex route-searching algorithm still runs on the remote cloud server, while GreenGPS is based on a global infrastructure and relies on high communication cost on uploading driving information. The experimental result shows that the relative error of our fuel consumption prediction results is close to GreenGPS by 10.7%-13.2% to 3.1%-4.9% for 6km to 32km testing-trips.
A. Estimating the Fuel Consumption Creating an FCM is not unique because it can be constructed from different types of data. Some work, such as [2], utilized vehicle speed to train a general fuel consumption computing equation, while [6] used both the vehicle speed and some other parameters of the vehicle, such as the vehicle type, the engine size, the car make, the loading and so on, to train the model. With the ubiquitous availability of sensors, such as GPS and OBD being standard in vehicles, many types of data can be easily collected, and driving behaviors can be inferred [1, 3, 5]. For example, Beusen et al. [5] summarized the main rules of fuel-efficient driving, e.g., maintain a steady speed by anticipating traffic flow and shut down the engine for longer stops. Then, good driving behavior can be recommended to drivers in order to decrease the fuel consumption [3, 7-9, 13-15]. Our work differs from previous work because: i) our FCM is locally constructed by the smartphone instead of by the remote server/cloud through uploading a large number of samples; ii) our FCM is personalized which is corresponding to a drivervehicle pair rather than a vehicle; and iii) we train the FCM based on OBD data stream in real time in order to adjust the model according various driving conditions.
EasyRoute can be used for different vehicles without knowing parameters of these vehicles. EasyRoute only requires the speed, the travel distance and the fuel consumption from OBD adaptor, which evolve along with factors that impact the fuel consumption, e.g., the experience of the driver, or the condition, the type and the loading of a vehicle. EasyRoute do
There are two methods and measurements to compute the fuel consumption, such as time-based method [16] (e.g., the driving time (m) multiplies the average fuel consumption (liter/m)) and distance-based method [10, 15] (e.g., the driving distance (km) multiplies the average fuel consumption
2
not collect any private information of the driver (e.g., driving experience) and the vehicle (e.g., the vehicle make, year or model), so it can work even after it only has run on the driver’s smartphone a few minutes to collect training data. For example, EasyRoute can predict the fuel consumption of a van in real time when its loading changes and can also predict that of a taxi when the number of passengers changes. Therefore, EasyRoute can be used for different vehicles.
IV. METHODS A. Data Models Two abbreviations are explained here in order to clarify the method: i) AFM refers to the Average Fuel consumption while Moving and its unit is liter/km; ii) AFW refers to the Average Fuel consumption while Waiting and its unit is liter/minute. 1) Model of the Personal Fuel Consumption The proposed personal FCM consists of two parts: i) FCM while moving and ii) FCM while waiting.
III. THE FRAMEWORK OF EASYROUTE The EasyRoute application runs on the driver’s smartphone and connects via Bluetooth or WiFi to the in-vehicle OBD sensors via an OBD adaptor. As shown in Figure 1, the framework of EasyRoute system has two parts: (i) on the smartphone side, data from the OBD of the vehicle is used to construct the personal FCM and the GPS trajectory of the travel is used to learn the traffic condition of passing roads in real time; (ii) on the server side, the server can gather traffic conditions of almost all roads from a large number of collaborative EasyRoute users, and then it searches the greenest route to a trip according to prediction results of total fuel consumptions via different routes.
Vehicle
EasyRoute server
Smartphone
Driver
OSM Road network
The FCM while moving is a discrete function of speed to AFM. The mapping (vi, f(vi)) denotes that the average fuel consumption is f(vi) liter/km if the vehicle speed is [vi-7, vi) km/h. 7 is the best step that we trained based on our dataset. f(vi) is computed by Eq. (1), 𝑓(𝑣𝑖 ) = ̅̅̅̅ 𝑀𝑖 ′,
where 𝑀𝑖′ = {𝑚|(𝑚 − 𝑀𝑖 ) ≤ 0.1, 𝑚 ∈ 𝑀𝑖 } , and Mi contains AFM Samples (AFMS) when the vehicle’s moving speed is in [vi-7, vi), and 𝑣𝑖 − 𝑣𝑖−1 =7. The FCM while waiting is also a discrete function of duration to AFW. The mapping (ti, g(ti)) denotes that the AFW is g(ti) liter/minute if the vehicle waits for ti-10 to ti seconds before moving. 10 is the default step and if AFW Samples (AFWS) are very sparse, the step should be larger. g(ti) is computed according to historical OBD observations by Eq. (2),
OBD
Train the personal FCM GPS
𝑔(𝑡𝑖 ) = 𝑁𝑖 ′,
Extract traffic information
Crowdsourced Traffic Info Traffic Infomation
Search the most economical route
2) Model of the Traffic Condition By mapping the GPS trajectory on the road network and analyzing the vehicle speeds at different timestamps, EasyRoute extracts a series of driving information on different roads and further forms the traffic condition. Two kinds of driving information are utilized: i) Speed-Distance pairs and ii) Waiting durations. Speed-Distance pairs are denoted by SD={sd1, sd2, …} and sd is 2-tuple (v, d), where, for example, the pair sd1=(v1, d1) refers that a vehicle moves d1 kilometers at the speed v1 km/h. Each v is rounded down to a multiple of 7 (which is the step of modeling the fuel consumption while moving). Waiting durations are denoted by WD={wd1, wd2, …}, and wdi denotes that the vehicle is waiting or is idling for wdi seconds before moving. For example, the traffic condition of a road is (SD, WD) where SD={(50, 1.4), (60, 0.18), (50, 1.2)} and WD={30, 10}. Then, (50, 1.4) means that the vehicle will move 1.4 km at the speed 50 km/h on this road, and {30, 10} means the vehicle might wait for 30 seconds and 10 seconds respectively on this road.
… Vehicle
(2)
where 𝑁𝑖′ = {𝑛|(𝑛 − 𝑁𝑖 ) ≤ 0.1, 𝑛 ∈ 𝑁𝑖 } and Ni contains AFWSes when the vehicle’s waits for [ti-10, ti) seconds, and 𝑡𝑖 − 𝑡𝑖−1 =10.
OSM Road network
Vehicle
(1)
Vehicle
Fig. 1. The framework of EasyRoute. The workflow of EasyRoute system can be summarized as: i) the EasyRoute application on the smartphone is activated when the smartphone receives the OBD data after the vehicle’s engine starts; ii) the smartphone efficiently trains the personal FCM after the vehicle moves a few minutes. The FCM along with a trip’s start point and destination is then uploaded to the EasyRoute server; iii) the EasyRoute server predicts total fuel consumptions of different routes according to the driver’s FCM and crowdsensing traffic information of roads; iv) the EasyRoute server ranks routes according to total fuel consumptions in ascending order and sends the top one to the driver.
B. The Method to Predict the Total Fuel Consumption The total fuel consumption of a route is calculated by summing the total fuel consumption on each road segment of
3
is mainly related to the driver’s driving habits (e.g., running/stop engine, using P gear or N gear). In order to collect sufficient AFWSes, the smartphone removes the past AFMSes when the trip begins, while it always keeps AFWSes since EasyRoute application was installed.
this route. The total fuel consumption of driving on a road segment is calculated by Eq. (3), |𝑊𝐷| ∑|𝑆𝐷| 𝑖=1 𝑓(𝑠𝑑𝑖 . 𝑣) ∗ 𝑠𝑑𝑖 . 𝑑 + ∑𝑗=1 𝑔(𝑤𝑑𝑗 ) ∗
𝑤𝑑𝑗 60
,
(3)
where (SD, WD) is the traffic condition of this road segment, and (f(), g()) is the driver’s FCM. C. An example of Training the FCM Given three drivers {d1, d2, d3}, their cars’ average fuel consumptions are 7.2, 9.1 and 7.8 liter/100km respectively. AFMSes and AFWSes from drivers {d1, d2, d3} are shown in Figure 2 (a) and (b) respectively. Each x mark represents an AFMS/AFWS and the darkness degree of the mark reflects the number of points at the same coordinate. As shown in Figure 2, on one hand, when the car’s speed increases, marks in the same column become concentrated; on the other hand, the average fuel consumption of both d1’s car and d3’s car are lower than d2’s car, which can be reflected by Average curves.
(a) Function f(v)
(b) Function g(t)
Fig. 3. The FCM of the driver d3 in Figure 2. D. Searching the Greenest Route The road network can be considered as a graph, whose edges are road segments and vertexes are intersections. The edge weight is the total fuel consumption of a vehicle driving on this road segment, then finding the eco-route is equal to finding the minimal weight path. The road network of a city is very large, so in order to reduce the route-searching space, the total fuel consumption and time of driving on the shortest route will be two constraints. For example, if the total driving time of a route (no matter whether it reaches the destination point or not) is larger than the time constraint, this route is abandoned. As shown in Figure 4, candidate routes are found and the economical route from A to E is A-C-D-E. Although the total fuel consumption of using sub route A-B-C is equal to using sub route A-C, the total travel time of using sub route A-B-C is longer, so we will finally recommend the sub route A-C.
(a) 1200 AFMSes of each car.
2:3 Start A point
3:1
B
5:8
1:2
C
2:3
D
1:1
E
Destination
5:7 7:8 x : y x denotes the total travel time, y denotes the total fuel consumption
Fig. 4. The graph of candidate routes tagged with the total travel time and the total fuel consumption.
(b) 180 AFWSes of each car. Fig. 2. AFMSes and AFWSes of three car-driver pairs. The circle denotes the average value of data in the same column. Generally, given a series of raw ODB readings, the number of AFWSes is much less than that of AFMSes. As shown in Figure 2, we got at least 1200 AFMSes but only got about 180 AFWSes by using the same number of OBD readings. Therefore, AFWSes might not be plenty enough to train the FCM. As shown in Figure 3, the fuel consumption of driver d3 is illustrated, where there is one null value when t=50 (s). The fuel consumption while moving is strongly related to the loading or the vehicle itself, however, the fuel consumption while waiting
Fig. 5. CDF of average fuel consumptions of 559 drivers. 4
V. EVALUATION OF FUEL CONSUMPTION PREDICTION
where D denotes the driver set, and |D|=559 in this paper. 4) Baselines Two commonly-used methods are baselines:
In this section, we will compute the accuracy of the total fuel consumption predicted by EasyRoute and other baselines and show experimental results.
i) Mean fuel consumption (Mean). The Mean approach uses the average liter/100km computed from data of the vehicle to predict the total fuel consumption (liter);
A. Experimental Setup 1) The Dataset The OBD observations and GPS trajectories of 559 taxis were collected from the Shenzhou taxi company in September 2015 in Beijing. The data collection frequency is 0.1 Hz. The mean travel distance and the mean travel time for each taxi was 177.8 km and 14.59 hours respectively. The CDF of average fuel consumptions of 559 taxis is shown in Figure 5. We can find that some taxis consume much more fuel than usual due to the terrible traffic jams in Beijing.
ii) Constant fuel consumption (Const). The Const approach uses 8.5 liter/100km (which is best according to training results) to predict the total fuel consumption. B. Experimantal Results Both the scale of training set (i.e., TrN) and that of testing set (i.e., TeN) might impact on the accuracy of the total fuel consumption prediction, so we change only one of them to evaluate our method.
2) Sampling The AFMSes can be two variants: i) using the instant speed at the sampling moment and ii) using the average speed in one sampling round. The experimental result shown in Figure 6 shows that ARE decreases when TeN increases. Meanwhile, ARE of using the instant speed for AFMSes is smaller than that of using the average speed, so we use the instant speed to sample AFMSes in the following experiments.
1) Using varied TeN and stataic TrN Experimental results shown in Figure 7 indicates that our method is the most accurate compared to the baselines. In order to further compare the accuracy of EasyRoute with baseline methods, we calculate the variance of REs of different drivers. As shown in Figure 8, the variance of REs of EasyRoute is still the smallest, which indicates that our method is more stable than the others.
Fig. 7. The accuracy comparison of different methods when TrN=1000 and TeN changes. Here, the minimal ARE of EasyRoute is 7.73%.
Fig. 6. Impacts on ARE brought by varying TeN and different speed measurements. Here TrN=1000. 3) Metrics We use different sizes of training dataset and testing dataset to evaluate our fuel consumption prediction method. The dataset of 559 drivers introduced in Figure 5 is used, and the data of each driver are divided into a trip for training and many trips for testing. TrN and TeN denote the number of AFMS/AFWS for training and for testing respectively. The testing sample set follows the training sample set and they are not intersected. Given a driver di ∈ D, Efci denotes the estimated total fuel consumption of a trip, and Gfci denotes the ground truth of the total fuel consumption of this trip. The Relative Error (RE) is computed by Eq. (4), and the Average Relative Error (ARE) of all drivers is computed by Eq. (5). 𝑅𝐸𝑖 =
|𝐸𝑓𝑐𝑖 −𝐺𝑓𝑐𝑖 |
𝐴𝑅𝐸 =
𝐺𝑓𝑐𝑖
∗ 100,
Fig. 8. Experimental results of variance of REs when TrN=1000 and TeN changes.
(4)
|𝐷|
∑𝑖=1 𝑅𝐸𝑖 |𝐷|
,
(5)
5
2) Using varied TrN and stataic TeN We aim to efficiently construct FCMs. As shown in Figure 9, we compare ARE of using EasyRoute with other methods when TrN varies. Experimental results show that EasyRoute can always achieve the lowest ARE even when TrN is only 50 (i.e., 500 seconds are cost to read OBD). Detailed evaluation results of EasyRoute are shown in Figure 10. When either TeN or TrN increases, ARE decreases. TeN and TrN can reflect the time length of data sampling for training and testing, and we will present experimental results based on the distance of driving next.
Fig. 12. Comparing different prediction methods with the average relative error when the length of trips-fortesting varies. The length of trips-for-training is 50 km. As shown in Figure 11, experimental results show that EasyRoute is the best one. The minimal error of the Mean approach is almost twice as that of EasyRoute. According to experimental results of GreenGPS [8], its relative error is about half of Mean approach too. Therefore, as a system that must efficiently predict the total fuel consumption of a trip, EasyRoute works well. As shown in Figure 12, experimental results show that relative errors of these three methods decreases much when the length of trip-for-testing increases. When the length increase from 10 km to 50 km, because more AFWSes can be got, the error of EasyRoute decreases from 14% to 10%.
Fig. 9. The accuracy comparison of different methods when TrN changes and TeN=100.
C. Discussion and Future Work First, the distribution of AFMS/AFWS of different drivers varies, this implies that parameters used by EasyRoute can be further personally trained. Second, the search cost for economical route recommendation is high, and we intend to develop heuristics to limit the search space and optimize the search time. Third, we do not consider the impact brought by slope roads on the fuel consumption in this paper because most roads in Beijing are flat except some ramps. We can refer to [8] to address this problem in the future work. As for future work, we will explore fine-grained driving behaviors and improve the FCM. Furthermore, we will deploy the system and make a longterm study to get feedback from drivers to improve its performance.
Fig. 10. ARE of using EasyRoute when TrN changes and TeN does not change. 3) Using varied distance of driving
VI. CONCLUSION We have presented EasyRoute, a personalized economical route recommendation system through the utilization of onboard diagnostic sensors installed in the vehicles. Personal FCMs are proposed and utilized to estimate the total fuel consumption while driving on different routes, and rowdsourced traffic information is utilized to recommend economical routes. Experimental results indicate that the personal FCM of EasyRoute can be quickly trained without requiring privacy information of drivers and vehicles in real time, upon which the relative error of the fuel consumption estimation method is around 10%. As for future work, we will explore fine-grained driving behaviors and improve the FCM. Further, we will deploy the system and make a long-term study to get feedback from drivers to improve its performance.
Fig. 11. Comparing different prediction methods with the average relative error when the length of trips-fortraining varies. The length of trips-for-testing is 10 km.
6
ACKNOWLEDGMENT
[8]
This work was partially supported by the National Basic Research Program of China (No.2015CB352400), the National Natural Science Foundation of China (No. 61602230, 61332005, 61373119).
[9]
REFERENCES [1]
[2]
[3]
[4]
[5]
[6]
[7]
[10]
K. Ahn, H. Rakha, A. Trani and M. Aerde. “Estimating vehicle fuel consumption and emissions based on instantaneous speed and acceleration levels”. Journal of transportation engineering, 2002, vol. 128, no. 2, pp. 182–190. T. Chan, Z. Ning, C. Leung, C Cheung, and et al. “On-road remote sensing of petrol vehicle emissions measurement and emission factors estimation in Hong Kong”. Atmospheric Environment, vol. 38, no. 14, 2004, pp. 2055–2066. J. Qian and R. Eglese. “Fuel emissions optimization in vehicle routing problems with time-varying speeds”. European Journal of Operational Research, vol. 248, no. 3, 2016, pp. 840–848. J. Shang, Y. Zheng, W. Tong, E. Chang and Y. Yu. “Inferring gas consumption and pollution emission of vehicles throughout a city”. Proc. of the ACM SIGKDD international conference on Knowledge discovery and data mining (KDD’14). ACM, 2014, pp. 1027–1036. B. Beusen, S. Broekx, T. Denys, C. Beckx, and et al. “Using on-board logging devices to study the longer-term impact of an eco-driving course”. Transportation research part D: transport and environment, vol. 14, no. 7, 2009, pp. 514–520. C. Guo, B. Yang, O. Andersen, C. S Jensen, and K. Torp. “Ecomark 2.0: empowering eco-routing with vehicular environmental models and actual vehicle fuel consumption data”. GeoInformatica, vol. 19, no. 3, 2015, pp. 567–599. P. Themann, J. Bock and L. Eckstein. “Optimisation of energy efficiency based on average driving behaviour and driver’s preferences for automated driving”. Intelligent Transport Systems, IET, vol. 9, no. 1, 2014, pp. 50–58.
[11]
[12]
[13]
[14]
[15]
[16]
[17]
7
X. Hu, V. Leung, K. Li, E. Kong, and et al. “Social drive: a crowdsourcing-based vehicular social networking system for green transportation”. Proc. of ACM international symposium on Design and analysis of intelligent vehicular networks and applications (DIVANet’13) , 2013, pp. 85–92. L. Kang, B. Qi, D. Janecek and S. Banerjee. “EcoDrive: A Mobile Sensing and Control System for Fuel Efficient Driving”. Proc. of ACM International Conference on Mobile Computing and Networking (MobiCom’15). ACM, 2015, pp. 358–371. O. Andersen, C. S Jensen, K. Torp and B. Yang. “EcoTour: reducing the environmental footprint of vehicles using eco-routes”. Proc. of IEEE International Conference on Mobile Data Management (MDM’13) , 2013, pp. 338–340. F. Saremi, O. Fatemieh, H. Ahmadi, and et al. “Experiences with greengps–fuel-efficient navigation using participatory sensing”. IEEE Transactions on Mobile Computing, vol. 15, no. 3, 2015, pp. 672–689. B. Guo, Z. Wang, Z. Yu, Y. Wang, and et al. “Mobile crowd sensing and computing: The review of an emerging human-powered sensing paradigm”. ACM Computing Surveys (CSUR), vol. 48, no. 1, 2015, artical 7. E. Yao and Y. Song. “Study on eco-route planning algorithm and environmental impact assessment”. Journal of Intelligent Transportation Systems, vol. 17, no. 1, 2013, pp. 42–53. I. Vlieger, D. Keukeleere and J. Kretzschmar. “Environmental effects of driving behaviour and congestion related to passenger cars”. Atmospheric Environment, vol. 34, no. 27, 2000, pp. 4649–4655. J. Meseguer, C. Calafate, J. Carlos Cano and P. Manzoni. “Assessing the impact of driving behavior on instantaneous fuel consumption”. Proc. of IEEE Consumer Communications and Networking Conference (CCNC’15), IEEE, 2015, pp. 443–448. J. Zhang, Y. Zhao, W. Xue and J. Li. “Vehicle routing problem with fuel consumption and carbon emission”. International Journal of Production Economics, vol. 170, 2015, pp. 234–242. Y. Wang, J. Jiang and T. Mu. “Context-aware and energy-driven route optimization for fully electric vehicles via crowdsourcing”. IEEE Transactions on Intelligent Transportation Systems, vol. 14, no. 3, 2013, pp. 1331–1345.
