The research is dedicated to solving the problem of people and vehicle localization in video frames. Video frames of areas with forest and roads are used as test data. The algorithm from the modified "deep_sort_realtime" package is used for object tracking. In addition, the capability to use Yolo 8 for object detection has been added, as well as the ability to extract informative features using Mobilenet v3. For the input images is used letterbox preprocessing, and various optimizations affecting the quality and speed of results have been added. For license plate recognition, the "tflite_avto_num_recognition" software package is used (which employs Canny and Hough transformations, as well as the CNN-LSTM-CTC neural network). The obtained solutions work in real time and rely on open-source libraries.
Download the conferece proceedings (PDF) from the official website: https://2024.dccn.ru/downloads/DCCN_2024_proceedings.pdf
Download the article (PDF) at the Intelligent Control Laboratory of the Program Systems Institute of the RAS website: https://icontrol.psiras.ru/publications/
Watch Vitaly Fralenko's presentation at Yandex Disk (from 42:30): https://disk.yandex.ru/i/cXOKWESjb5szYg
Vitaly Fralenko, Mikhail Khachumov. A practical solution to the problem of detecting peoples and vehicles from video frames // Distributed computer and communication networks: control, computation, communications (DCCN-2024) : Proceedings of the XXVII International Scientific Conference. Russia, Moscow, September 23–27, 2024. Moscow : Peoples’ Friendship University of Russia Named After Patrice Lumumba, 2024. Pp. 45–50.