Pedestrian Detection and Vehicle Classification using Deep Convolutional Neural Networks for Video Surveillance Applications

Today, we come across examples of artificial intelligence in many areas of our lives with the development of machine learning and deep learning networks. One of these areas is to combine surveillance cameras with artificial intelligence to perform a determined task without a personnel. To make its job description more understandable, we can say that using artificial intelligence to detect and classify objects in real-time surveillance camera images. Although this task may seem easy for us, it is extremely important to perform the detection task quickly and accurately when performed on the machine side. In this direction, an end-to-end system solution has been developed by using deep learning networks for the detection of intruders in 8 classes including people and vehicles in outdoor live surveillance cameras. First, the state-of-the-art technology algorithms capable of fast and accurate detection that can be used for object detection and classification task to be performed on the end device provided by Videotec company were examined and it was decided to use YOLO version 4 in the preliminary tests. The YOLO model is used with pre-trained weights to reduce time and power costs instead of being trained from scratch. In order to test the performance of the model in harsh conditions, a dataset in MSCOCO standard was created, in which images are taken from live surveillance camera images in addition to samples taken from known datasets such as MSCOCO, iLiDS, UA-DETRAC. The dataset that we call critical conditions includes various scenarios such as black/white, occlusion and bad weather. Afterwards, various tests were carried out using Tensorflow official tools and the values ​​such as inference time and footprint of the model were compared with different models and tested on the specified hardware. Furthermore, some of the real-time tests were performed on Google Colab due to the lack of hardware and the results were recorded. As a result of all these tests, it is seen that the lack of GPU usage determined within the scope of the project is the biggest bottleneck and it is determined that the CPU was insufficient.