Traffic State Detection Using Acoustics Nitakshi Sood, Arshvir Kaur, Bhavdeep Sachdeva, Dinesh Vij, Naveen Aggarwal DIC, UIET, Panjab University, Chandigarh Introduction Initial Findings Results Train set = 161 recordings per class Test set = 69 recordings per class Testing repeated for 6 times The algorithm below describes the binary classification techniques applied. The table shows the classification accuracy (in %) of the traffic classes. Traffic Scenes Considered • BusyStreet: Heavy flow traffic means the congested road or the jammed situation. Vehicle speed between 0-20 kmph • MediumStreet: traffic state in which volume of vehicles is moderate. Vehicle speed between 20-50 kmph • QuietStreet : This is the traffic scene with little traffic. Vehicle speed between more than 50 kmph Objectives : The traffic and number of vehicles on roads are increasing with an unstoppable pace, which in turn leads to the problem of traffic congestion. We propose the use of Acoustics to determine : (a) How efficient and accurate do acoustic signals turn out to be in the detection of Traffic events. (b) If acoustics could be used as a standalone parameter to detect Traffic events. Traffic state Acoustic Analysis – What and Why? Data Collection • Crowd sourced data collected from 7 different Smart Phones using “Smart Voice Recorder by SmartMobv1.71” application. • Phones used: Nexus, Samsung S3/S4/Grand, lenovo,Mi4. • 1000 recordings of length 30 seconds approximately - for each traffic scene, collected in almost 15 sessions. • Recorded at 16 KHz Sample Rate and 16-bit Bit Rate. • Additional data captured such as location of the phone, whether the phone was in hand/pocket/bag, associated AndroSensor file, number of microphones in the phone that was used for recording, air (still or turbulent), and environmental conditions. • Acoustic Event: Acoustic events represents all the events that are present in audio signals; such as engine start, engine stop, horn, music, speech, etc. • Acoustic Scene: These events usually occur within different acoustic scenes such as car, bus, train, busy street, quiet street, etc. What can we get on capturing an Acoustic signal? Advantages of using acoustics: Relatively inexpensive sensors. Less energy consumption. Not affected by visual occlusions/orientation/lighting conditions. Works for non-lane driven traffic. Prominent at slower speeds Prominent at faster speeds Others Honks Tire noise Environmental noise Engine-idling noise Engine noise Exhaust noise Air Turbulence Classifier Congestion vs Non- congestion Medium vs Free flow Overall accuracy GMM 77.8% 61.6% 69.7% SVM 92.7% 94.2% 93.2% Neural Network 90% 93.5% 91.75% Classes Number of recordings Heavy flow 280 Medium flow Free flow Quiet street 160 Two level classification greatly improves the overall classification accuracy. Also the results show that SVM and NN out perform GMM. Conclusion • Extensive data collection: inside pocket, inside bag, windy conditions, rainy season, on different vehicles is required for future. •Two level classification done with Hamming window function and frame size of 8192 with suitable feature set of (MFCC+SC+EE) gives the best results for traffic state detection. Features • Temporal features : Zero Crossing Rate (ZCR) Short Time Energy (STE) Root Mean Square(RMS) Energy Entropy (EE) • Spectral features : Spectral Centroid (SC) Spectral Roll Off Point (SRP) Spectral Flux (SF) Mel Frequency Ceptral Coefficients (MFCC) Process of Traffic State Detection using Acoustics Future Work • Early Detection: Minimum amount of recording needed by the classifier to correctly identify a scene. • Triggering Mechanism: Audio would not be recorded continuously. Some event such as accelerometer would trigger the microphone of smartphone. • Convulational Neural Network Implementation: To implement Deep learning to produce efficient accuracy. Experiments on Frame size and Window Type Our approach Rather than using Infrastructure Based Solutions (installing sensors into vehicles and road ways)  Crowd sourced data from Smart Phones using app. To make it efficient, experiments need to be done on: • Extensive Data Collection • Window Size & Type • Early Detection • Feature Set • Classification Technique • Role of Learning • The audio signal captured was segmented using small windows. • Therefore, primary analysis window (Hamming Window) of various sizes as illustrated were tested in this work. • Classifier used: GMM, SVM, Neural Network Contact details For further information contact us at : - navagg@gmail.com - vijdinesh@gmail.com - bhavdeep.mailme@gmail.com - nitakshi.sood@gmail.com arshviranttal77@gmail.com Acknowledgement We would like to thank the Design Innovation Centre (DIC), UIET for extending support in various forms.