Literature Review

Classification of Road Surface Disruptions – Literature Review

Paper Key
Paper 1 (P1)	(2017) Luis C. González et al. Learning Roadway Surface Disruption Patterns Using the Bag of Words Representation
Paper 2 (P2)	(2011) Kongyang Chen et al. Road Condition Monitoring Using On-board Three-axis Accelerometer and GPS Sensor
Paper 3 (P3)	(2008) Jakob Eriksson et al. The Pothole Patrol
Paper 4 (P4)	(2008) Prashanth Mohan et al. Nericell
Paper 5 (P5)	(2011) Mikko Perttunen et al. Distributed Road Surface Condition Monitoring Using Mobile Phones

Keywords:
Road surface disruptions	embedded inertial sensors	Road condition monitoring	International Roughness Index
Mobile sensor computing system	Intelligent transportation systems	Geometric mean of accuracies	Relative sensitivity

Problems identified by the paper:

      P1.

1. No publicly available data sets to compare and contrast RSD approaches.
2. Location of sensor inside/outside the vehicle has been constrained preventing seamless and transparent technology.
3. Poor feature vectors have been developed so far.

     P2.

1. GPS data readings are often missed because the sensor stops working among tall buildings and tunnels
2. GPS receiver gives invalid points right after it is just turned on.
3. Data is not normalized and does not follow SI units and standards.     

P3.

1. Road surface monitoring is a problem that requires mobility to solve and static sensors cannot prove effective.
2. There are a large number of benign events such as door slamming, braking, sudden serves that produce high energy acceleration signals
3. Potholes that need serious fixing need to be distinguished from minor bumps on the road that do not need urgent attention.
4. Effects experienced due to an anomaly depend on how that particular road anaomaly was approached, with what speed and under which orientation of the sensors.
5. Energy content of the signal produced cannot be an accurate estimate of which class the event belongs to.

     P4.

1. Most of the work in road surface monitoring has been done in developed countries, whereas road and traffic conditions in developing countries are more varied, heterogeneous and chaotic.
2. Accelerometer inside the smartphone needs to be virtually reoriented to make up for the arbitrary orientation of the phone it is embedded in.

     P5.

1. Speed dependence: driving over the same road surface anomaly with different speeds results in different signal patterns.
2. GPS signal from phone is very noisy and raw data shows sudden bursts due to the process switches and priorities inside the phone.
3. Accuracy alone is not a reliable measure t model goodness in case of skewed data set.

Solutions proposed by the paper:

P1.

1. New and more heterogeneous data set.
2. Selects a better feature vector that enhances classification performance.
3. Reports the best accuracy results obtained by an algorithm for this problem.      

P2.

1. Data cleaning algorithm: delete “bad zones” with missing GPS values and outliers. Interpolate the remaining data points at an interval of 0.01s.
2. Normalize the acceleration between range -2g to 2g and converts velocity to km/h units
3. Develops a pavement roughness level algorithm that classifies a road on the basis of International Roughness Index (IRI)

     P3.

1. Uses inherent mobility of cars to monitor and assess road surface conditions and generate a carefully hand-labeled data.
2. Use peak x and z acceleration along with the velocity to minimize false positives (0.2%).
3. Use loosely labeled data set in which they know a rough number of road anomalies but not their exact locations.
4. Cluster detections based on the number of times they occurred to filter out spurious events.

     P4.

1. Nericell is a system for rich monitoring of road and traffic conditions using smartphones in the city of Bangalore.
2. Developed an algorithm to reorient a disoriented accelerometer inside a phone.
3. Braking event can be distinguished from a pothole occurrence in a sense that the former last for relatively long time frames whereas the latter last for only a fraction of a second.

    P5.

1. Robust feature extraction technique which removes speed dependency.
2. Preprocesses the data using GPS outlier removal and Kalman filter to further reduce the noise.

Dataset

• Route, Hardware & Vehicles:

P1.

1. 130 km Chihuahua, Mexico.
2. 2013 Motorola Moto G smartphones powered by a 1.2GHz processor with Android OS version 4.4.4., and tri-axial accelerometer ST Micro LIS3DH. 50 Hz sampling rate for the accelerometer.
3. 12 cars and trucks. Tried six different locations for the smartphone inside the vehicle.

P2.

1. Urban area of Beijing.
2. MEMS motion module, LIS33DE. 400 kHz read/write clock frequency and 100Hz sampling frequency of accelerometer. GPS receiver is the vehicle-mounted HM-CZ02 GPS sensor with 1Hz frequency. Client microprocessor is a 32 bits high performance Cortex ARM chip STM32F103
3. Accelerometer mounted on right side of the dashboard. GPS on left-front windshield in the lateral part.   

                 P3.

1. 2492 distinct km of roads in greater Boston metropolitan and 9730 km in total.
2. A Soekris 4801 embedded computer running Linux,a WiFi card, a Sprint EVDO Rev A network card, an external GPS with 1Hz sampling frequency (mounted on the roof of the car), and a 3-axis accelerometer with 380Hz sampling frequency (mounted in the same place (dashboard) and with the same orientation in each car)
3. 7 cabs (Toyota Priuses) deployed over a period of 10 days.

                 P4.

1. Bangalore.
2. Windows Phone 5.0. (HP Ipaq HW6965, HTC Typhoon) Sparkfun WiTilt Accelerometer with 310 Hz sampling frequency. Accelerometer placed in back and middle seats, dashboard and hand rest of the vehicle.
3. Toyota Qualis multi-utility vehicle, 4 weeks drive 62km in total.

                 P5.

1. Urban setting (Country/city not mentioned).
2. Nokia N95 8GB. Accelerometer samples at 38Hz and GPS at 1Hz. Attached to a rack on the windshield of the vehicle.
3. Single passenger car, 40 minutes’ drive 25km in total.

• Raw Sensing data variables

P1.

< GPS point, z-axis acceleration >

P2.

< time, location, velocity, three-axis acceleration >

P3.

< time, location, speed, heading, three-axis acceleration >

P4.

< GPS, x-z acceleration >

P5.

< time, latitude, longitude, three-axis acceleration >

• Number of instances and target classes

                   P1.  500 instances. 5 categories:

1. Potholes
2. metal bumps
3. asphalt bumps
4. worn out road
5. regular road

                   P2.  Classifies pavement roughness into four classes:

1. Excellent
2. Good
3. Qualified
4. Unqualified

P3.  Three sets of data: Hand labeled, loosely labeled and uncontrolled cab data. Event classes were:

1. Smooth road
2. Crosswalks/Expansion joints
3. Railroad crossing
4. Potholes
5. Manholes
6. Hard stop
7. Turn

P5.  Data was labelled manually with the help of the recorded video of the drive. Cobblestone segments were discarded from the data using road database. The anomalies were categorized into two classes:

1. Type1 represents small potholes, rail road crossing and other road surface roughness. 184 in total, lasting for 613.9 s (~ 10 minutes)
2. Type 2 represents man-made speed bumps and severe anomalies that can cause accidents at high speeds. 42 in total, lasting for 81.6 s (~ 1 minute)

• Cross-Validation/Percentage Split

                     P1. 60% training, 40% testing. Repeated 30 times randomly.

                     P2. Not mentioned

         P3. Random split

                     P4. Not mentioned

   P5. 5-fold cross validation. Training set of each fold contains 4/5 of total anomaly      windows and 4/5 of normal windows. Folds were created from consecutive windows.

Machine Learning techniques used by the paper:

P1.

Bag of Words approach for preprocessing. 7 classifiers for classification (first six were implemented in Python using scikit-learn package whereas KR was coded in MATLAB using CLOP):

1. Artificial Neural Networks ANN)
2. Support Vector Machines (SVM)
3. Decision Trees (DT)
4. Random Forrest (RF)
5. Naive Bayes Classifier (NB)
6. K-nearest
7. neighbors (KNN)
8. Kernel Ridge (KR)

     P2.

Data is preprocessed by removing outliers and missing values. Data points are interpolated using 10ms interval. Power Spectral Density is calculated using auto correlation function.

     P3.  

1. The time series is segmented into 256 sample windows.
2. Windows in which the speed is nearly zero or very low, are discarded.
3. The window is then passed through a high pass filter to remove low frequency components.
4. Windows with z-acceleration peak less than a certain threshold tzare rejected.
5. Windows with x-z ratio less than tx are discarded.
6. Windows with speed-z ratio less than ts are discarded.
7. The three thresholds/tuning parameters are found using combined exhaustive search.

      P4.  

1. The mean of x-acceleration is calculated over a sliding window of N seconds wide. If the mean exceeds a threshold T, it is declared as a braking event.
2. C# and python
3. Proposes two bump detectors based on speed. At high speed (> 25kmph) only z-peak greater than a threshold T (1.75g) is detected. At low speed sustained bump in which az falls below threshold T (0.8g), lasting 20 ms (~7 samples) is observed.

      P5.  

1. Data collection tool was written in JAVA.
2. Data was framed using a sliding window using multiple frame lengths from 0.5s to 2s.
3. Feature extraction was done using a sliding window of 2s and slide being 0.5s
4. Following features for each dimension were extracted for each window:
1. Standard deviation
2. Mean
3. Variance
4. Peak-to-Peak
5. Signal magnitude area
6. 3-order autoregressive coefficient
7. Tilt angles
8. Root mean square
5. Autocorrelation between signals of all dimensions were also calculated because x-y-z signals showed similarity in case of anomalies (except a pothole)
6. FFT energy was extracted from 17 bands for each direction and mel-frequency cepstral coefficients in 4 bands.
7. Backward feature selection algorithm of PRTools was used to select the most optimum set of features.  
8. SVM with soft margin and radial basis kernel was used as a classifier.
9. Misclassification weights were set according to the number of normal and anomaly instances to deal with skewed data.

Evaluation Metrics:

P1
a. Accuracy = TP+TNTP+TN+FP+FN
b. Area under ROC curve

     P2.

1. Standard deviation of pavement roughness
2. International Roughness Index
3. Riding Quality Index (metric for driving comfort)

     P3.

1. (Hand labeled data):

Detector score = correct_detections – incorrect_detections2

2. (Hand labeled + Loosely labeled data):

Detector score = corr – labelled_incorr2 - max⁡(0,incorrloose- countr

     P4.        False positives and false negatives

     P5.  

1. Geometric mean= sensitivity*specificity
2. Relative sensitivity= sensitivityspecificity

       

Weaknesses of the Paper

P2.

1. Removes large chunks of data which is not transmitted properly or has missing GPS values. Could use better strategy for dealing with missing values.
2. Does not give supporting evidence for why the acceleration values were normalized between -2g to 2g range (where g = 9.8m/s2).  
3. The threshold on the basis of which the outliers were removed is not explicitly defined.
4. The scope of the paper is limited as it uses Chinese industry standards to classify pavement roughness levels.
5. No evidence/reference has been cited that supports why pavement roughness can be modelled as a Gaussian distribution with zero mean.
6. Number of instances and nature of vehicles have not been explicitly mentioned.

P3.

1. The data is not representative of true frequencies of road surface condition (Smooth road is prevalent in real life but reduced in the data set to prevent it from being skewed).
2. Hand labeled data has few instances for non-pothole events which results in over detection of potholes in an unseen dataset.   

P4.

1. The accelerometer readings have been recorded in a highly controlled environment, under fixed speeds which is far from reality and results in overfitting.
2. It uses an external database to post-process intended bumps (speed breakers etc.) from unintended ones (potholes). This limits the scope of the work to countries which have a well-maintained road database.

P5.

1. The process of estimating speed from GPS points seems weak. Speed is reduced to zero where acceleration was zero without any supporting evidence that the acceleration readings were truly related to the velocity of the vehicle. Furthermore, the correctness of the speed is checked from the video which is not a very strong method.