Signal Strength and Sensing System Configuration Velocity Estimation Right-turn Detection Vehicle Classification • AMR-3 is placed 0.9 meters from AMR-1 and these two sensors are used for vehicle velocity estimation. • The relatively short distance between the sensors removes the problem due to a vehicle performing a maneuver which may be detected only by one of the two sensors. • A reliable method to calculate the time delay between the two sensor locations is by using the cross-correlation between the sensors signals. • The time delay in terms of samples is given by = arg max = 1 3 − −1 =0 − −1 ≤≤−1 • DSP techniques are used to reduce the computational effort. • A test vehicle equipped with carrier-phase GPS has been used to verify the accuracy of the proposed velocity estimation method. • The velocity estimates are multiplied by the factor = min( −1 −3 , −3 −1 ) to account for misalignment of the sensors and get zero-offset estimates. Portable Roadside Sensors for Vehicle Counting, Classification and Speed Measurement Saber Taghvaeeyan and Rajesh Rajamani Introduction Robustness to Traffic on the Non-adjacent Lane • A portable sensor system is designed that can be placed adjacent to the road and can be used for vehicle counting, speed measurements and vehicle classification. • The sensor system consists of magnetoresisitve devices that measure magnetic field and associated signal processing algorithms. • The sensor system can make these traffic measurements reliably for traffic in the lane adjacent to the sensors. The vehicle detection rate accuracy is 99%. • The developed signal processing algorithms enable the sensor to be robust to the presence of traffic in other lanes of the road. • The velocity estimation has a max error of 2.5% over the entire speed range 5 – 60 mph. • Vehicle classification is done based on the magnetic length and an estimate of the average vertical magnetic height of the vehicle. • The sensor system can be used to reliably count the number of right-turns at an intersection. • The developed sensor system is compact, portable, wireless and inexpensive. • Signals from 216 vehicles driving in the non-adjacent lane were also recorded. • Passengers vehicles driving in the non-adjacent lane typically do not create detection errors. • However, larger vehicles (trucks, buses, etc.) in the non-adjacent lane may create large enough signals to cause over-detection and affect accuracy of the system. • 15 vehicles out of 216 vehicles created a large enough signal to be miscounted as vehicles passing in the adjacent lane. • If uncorrected, this will cause an over-detection error of 8%. • Similar error rates (7-15%) have been reported in literature even for magnetic sensors placed in the middle of the lane. 1 Use of AMR-2 to reject errors due to traffic passing in the non-adjacent lane • It is shown that the magnetic field intensity around a vehicle has a relation that approximately varies as 1/ with distance, where is the distance from the vehicle. 2 • Hence, the ratio 2 1 should be larger for vehicles in the non-adjacent lane, compared to vehicles passing in the adjacent lane. • Also the vehicles passing in the non-adjacent lane have a much lower peak value, , on average compared to vehicles passing in the adjacent lane. • These two metrics can be used to reject the traffic passing in the non-adjacent lane affecting the sensors. • The following figure shows the result of applying the proposed method to the data set. • A Support Vector Machine has been used to come up with the classification boundary. • Using the proposed method, the error reduces from 8% to 1%. 1 J. Medina, A. Hajbabaie and R. Benekohal. Detection performance of wireless magnetometers at signalized intersection and railroad grade crossing under various weather conditions. Transportation Research Record, pp. 233-241. 2011. 2 S. Taghvaeeyan and R. Rajamani. Use of vehicle magnetic signatures for position estimation. Applied Physics Letters 99(13), pp. 134101-134101-3. 2011 • Knowing the time duration and velocity of each passing vehicle, the magnetic length of the vehicle can be calculated and used for vehicle classification. • Vehicles are divided into four classes, Class 1: Sedans, Class 2: SUVs, Vans and Pickups, Class 3: Buses and 2,3-axle Trucks and Class 4: Articulated Buses and 4,5-axle Trucks. • Since vehicles in class I and class II have similar length and consequently similar magnetic lengths, it is not possible to classify them by using only magnetic length. • It is expected that magnetic component locations of a vehicle in Class II lead to a higher magnetic height compared to vehicles in Class I. • Placing another sensor, AMR-4, one foot vertically above AMR-1, it is expected that the ratio −4 −1 will be larger for vehicles in Class II. • This ratio along with the magnetic length can be used to determine boundaries for classifying Class I and Class II vehicles with an accuracy of 83%. • Using just one AMR sensor as shown, the number of right-turns at an intersection can be counted. During the experiments, 56 out of 59 right-turns were counted correctly resulting in a detection rate of 95%. • Typically straight-driving vehicles are not detected, since they pass at a larger distance from the sensor compared to vehicles making a right turn. • However larger straight-driving vehicles can create large enough signals to be miscounted as vehicles making right turns. • During the experiments, 18 straight driving vehicles created large enough signals to be miscounted as right-turning vehicles which results in a detection error of 31%, if uncorrected. • Two methods, A and B, are proposed to identify and reject the errors caused by straight driving vehicles, using two and four AMR sensors respectively. • Considering the shown sensor configuration, integrating the signals from 4 AMR sensors of each detected vehicle we expect the following • Method A: The ratio = −2 −3 should be closer to 1 for straight driving vehicles since they pass at larger distances from the sensors. • Method B: A plane is fit to the measurements from the four AMR sensor. By considering the angle of the plane, , the straight-driving vehicles can be excluded. • The two methods can be used separately or combined. With classification boundaries, straight- driving vehicles can be completely excluded reducing the 31% misdetection error to zero. : AMR Sensors 1 3 2 4 Scenario 1: Straight on Lane 1 Scenario 2: Right turn from Lane 1 to Lane 2 Scenario 3: Straight on Lane 2 d = 20 cm x y z 1 2 3 4 Scenario 1: −1 ≅ −3 > −2 ≅ −4 Scenario 2: −3 > −1 ≅ −4 > −2 Scenario 3: −3 ≅ −4 > −1 ≅ −2 • The 3-axis HMC2003 set of AMR devices from Honeywell are utilized. • The signal levels are typically 10 times smaller when sensors are placed adjacent to the road compared to the case when sensors are placed on-road in the center of the lane. • Sensors outputs are amplified to get better signal-to-noise ratio and for use of the signals for vehicle counting, speed measurement and classification. • The following figure shows the configuration of the sensing system. • AMR sensors 1 and 2 are used to obtain an estimate of lateral location of the vehicle. • AMR sensors 1 and 3 are used to calculate the longitudinal velocity of the vehicle. • AMR sensors 1 and 4 are used to get a rough estimate of the average vertical magnetic height of passing vehicles. Vehicle Detection and Counting • Magnetic readings of the Z axis of AMR-1 are used for detecting and counting the passing vehicles in the adjacent lane. • A threshold of 30 counts was used as the vehicle detection threshold. • Signals from 188 vehicles driving in the adjacent lane were recorded, 186 vehicles created a large enough signal to be detected resulting in a detection rate of 99%. 5.5 6 6.5 7 7.5 8 -150 -100 -50 0 50 100 150 200 250 300 350 Magnetic Field Readings - Ford Ranger - Sensor on the Road time (sec) B (counts) B x B y B z 9 9.5 10 10.5 11 11.5 12 12.5 13 -25 -20 -15 -10 -5 0 5 10 15 20 Magnetic Field Readings - Ford Ranger - Side of the Road time (sec) B (counts) B x B y B z 0 50 100 150 200 0.85 0.9 0.95 1 B 1-max (counts) B 2 /B 1 Adj Lane Non-adj Lane Class. Bound. 5 10 15 20 25 30 -15 -10 -5 0 5 10 Velocity Estimation Error GPS Velocity (m/s) Error (%) Threshold Method Cross-corr. Method Class I Class II Class III Class IV 0 5 10 15 20 25 Magnetic Length (m) 2 4 6 8 10 0.8 1 1.2 1.4 1.6 1.8 2 Magnetic length (m) B 4-z / B 1-z Class I Class II -20 0 20 40 60 80 100 120 55 60 65 70 75 80 85 90 95 (degrees) r (%) Right turn Straight on Lane 1 Straight on Lane 2 1 3 2 : AMR Sensors 1 2 3 90 cm 10 cm 4 4 AMR Sensors Side Walk Lane 1 Lane 2