In the previous chapter, you got your hands on deep convolutional neural networks and built deep classification and localization networks using transfer learning. You have started your deep learning journey and have familiarized yourself with a range of deep learning concepts. You now understand how deep models are trained and you are ready to learn about more advanced deep learning concepts.
In this chapter, you will continue your deep learning journey, first using object detection models to detect multiple objects of different types in a video of a relevant scene such as a street view with cars and people. After that, you will learn how such models are built and trained.
In general, robust object detection models have a wide range of applications nowadays. Those areas include but are not limited to medicine, robotics, surveillance, and many others. Understanding how they work will allow you to use them for building your own real-life applications, as well as elaborating on new models on top of them.
After we cover object detection, we will implement the Simple Online and Realtime Tracking (Sort) algorithm, which is able to robustly track detected objects throughout frames. During the implementation of the Sort algorithm, you will also get acquainted with the Kalman filter, which in general is an important algorithm when working with time series.
A combination of a good detector and tracker finds multiple applications in industrial problems. In this chapter, we'll limit the applications by counting the total objects by their type as they appear throughout the video of the relevant scene. Once you understand how this specific task is achieved, you will probably have your own usage ideas that will end up in your own applications.
For example, having a good object tracker allows you to answer statistical questions such as which part of the scene appears more condensed? And, where do objects move more slowly or quickly during the observation time? In some scenarios, you might be interested in monitoring the trajectories of specific objects, estimating their speed or the time that they spend in different areas of the scene. Having a good tracker is the solution for all of these things.
In this chapter, we will cover the following topics:
- Preparing the app
- Preparing the main script
- Detecting objects with SSD
- Understanding object detectors
- Tracking detected objects
- Implementing a Sort tracker
- Understanding the Kalman filter
- Seeing the app in action
Let's start the chapter by pointing out the technical requirements and planning the app.