View Independent Vehicle Model Recognition Using Semantic Segmentation and Image Retrieval

Vehicle model recognition benefits the parking, surveillance and tolling system by automatically identifying the exact make and model of the passing vehicles. Current work generally relies on a specific view of the vehicle or requires lots of images from various perspectives. This paper proposes a novel method based on semantic segmentation and image retrieval techniques, which enables the authors to recognize the vehicle model from an arbitrary and unknown view. More than 8,000 vehicle images are labeled with their components’ information, such as headlights, windows, wheels, logos, etc. A deep convolutional neural network (CNN) is developed to segment and localize the typical components. The vehicle postures are estimated by the spatial relationship between different segmented components. Due to the variety of the perspectives, a 7-bit vector is defined to represent the relative posture of the vehicle and screen out the images with similar photographic perspective. Two approaches are employed to extract the features from each image patch: (1) The speed up robust feature (SURF) combined with the bag-of-features (BoF) and (2) pre-trained CNN. A spectral saliency map is used as the mask of the original image to retain the essential features and filter the redundant points out. The descriptors of each component are then weighed and aggregated to retrieve the best matching image from the database. The results showed its accuracy (89.6%) and efficiency, demonstrating the vast potential of applying this method to large-scale vehicle model recognition.

• Supplemental Notes:
  • This paper was sponsored by TRB committee ABJ70 Standing Committee on Artificial Intelligence and Advanced Computing Applications.
• Corporate Authors:

  Transportation Research Board

  ,    
• Authors:
  • Liu, Chenglong
  • Deng, Fuwen
  • Jiang, Ying
  • Pu, Ziyuan
  • Du, Yuchuan
  • Wang, Yinhai
• Conference:
  • Transportation Research Board 98th Annual Meeting
  • Location: Washington DC, United States
  • Date: 2019-1-13 to 2019-1-17
• Date: 2019

Language

• English

Media Info

• Media Type: Digital/other
• Features: Figures; Photos; References; Tables;
• Pagination: 13p

Subject/Index Terms

• TRT Terms: Automatic vehicle classification; Image processing; Neural networks; Vehicle detectors
• Subject Areas: Highways; Operations and Traffic Management; Vehicles and Equipment;

Filing Info

• Accession Number: 01697456
• Record Type: Publication
• Report/Paper Numbers: 19-03448
• Files: TRIS, TRB, ATRI
• Created Date: Mar 1 2019 3:50PM