Vehicle-pedestrian classification with road context recognition using convolutional neural networks

• Main Authors: Billones, Robert Kerwin C., Bandala, Argel A., Gan Lim, Laurence A., Culaba, Alvin B., Vicerra, Ryan Rhay P., Sybingco, Edwin, Fillone, Alexis M., Dadios, Elmer P.
• Format: text
• Published: Animo Repository 2019
• Subjects: Vehicle detectors; Pedestrian inertial navigation systems; Context-aware computing; Sensor networks; Traffic flow; Manufacturing; Mechanical Engineering
• Online Access: https://animorepository.dlsu.edu.ph/faculty_research/2990
• Institution: De La Salle University

In road traffic scene analysis, it is important to observe vehicular traffic and how pedestrian foot traffic affects the over-all traffic situation. Road context is also significant in proper detection of vehicles and pedestrians. This paper presents a vehicle-pedestrian detection and classification system with road context recognition using convolutional neural networks. Using Catch-All traffic video data sets, the system was trained to identify vehicles and pedestrians in four different road conditions such as low-altitude view T-type road intersection (DSO), mid-altitude view bus stop area in day-time (DS4-1) and night-time (DS4-3) condition, and high-altitude view wide intersection (DS3-1). In the road context recognition, the system was first tasked to identify in which of the four road conditions the current traffic scene belongs. This is designed to ensure a high detection rate of vehicles and pedestrians in the mentioned road conditions. Road context recognition has 98.64% training accuracy with 2800 sample images, and 100% validation accuracy with 1200 sample images. After road context recognition, a detection algorithm for vehicle and pedestrians was trained for each condition. In DSO, the training accuracy is 97.75% with 1200 image samples, while validation accuracy is 94.75% with 400 image samples. In DS3-1, the training accuracy is 98.63% with 1400 image samples, while validation accuracy is 98.29% with 600 image samples. In DS4-1, the training accuracy is 99.43% with 1400 image samples, while validation accuracy is 99.83% with 600 image samples. In DS4-3, the training accuracy is 97.77% with 1400 image samples, while validation accuracy is 98.29% with 600 image samples. © 2018 IEEE.