Analysis of the vehicle's flow based on the neural network and the SIFT method

Abstract

The article presents a vehicle counting system based on TensorFlow neural network models and the SIFT machine vision method. An experimental comparison was made of five detectors consisting of metaarchitecture (Faster R-CNN, SSD) and neural networks extracting features (Resnet V1 100, Inception V2, Inception Resnet V2 and Mobilenet V1). The main aspects of these detectors are analyzed, such as accuracy, speed, memory consumption, the number of floating point operations per second and the number of trainable parameters of convolutional neural networks. The calculation of vehicles is carried out by an algorithm based on the SIFT method. This algorithm compares the descriptors of all vehicles in the frame at the current time with the descriptors at the previous time. Based on the maximum match of the descriptors, the algorithm assigns the vehicle identification number from the previous frame, and in the absence of matches creates a new identification number. This approach will make it possible to calculate vehicles more accurately and assess their trajectory and speed.