In this blog post, we’ll explore how algorithms and artificial intelligence technologies are used to detect vehicles and identify vehicle types in footage from traffic accidents, dashcams, and CCTV cameras.
Thousands of people die in traffic accidents in South Korea each year. According to the Traffic Accident Analysis System (TAAS) of the Korea Road Traffic Authority, the number of traffic accident fatalities in 2025 was recorded at 2,549. Although the number of traffic accident fatalities has decreased significantly compared to the past, traffic accidents—both major and minor—continue to occur regularly. As accidents continue to occur, the use of in-vehicle dashcams is becoming increasingly important for securing accurate footage of accident scenes. By utilizing algorithms that will continue to advance in the future, it will be possible to automate the analysis of vehicles involved in accidents or criminal acts captured by dashcams, thereby enabling faster post-incident processing of accidents and incidents. In this article, we will first explore what an algorithm is, and then examine the specific sequence and methods by which algorithms for vehicle analysis operate. Let’s start by understanding what an algorithm is. Programs operate based on algorithms. An algorithm is a set of procedures or a problem-solving method that specifies what result will be produced based on a given input. For example, when walking down a street and encountering a fork in the road, a person can decide which path to take based on the surrounding circumstances or their destination. In contrast, since computers do not understand situations or make judgments on their own like humans do, criteria and procedures for solving problems must be pre-designed or taught to the program. An algorithm is a set of rules that determines how to interpret and process the input information to produce a specific result. Now, let’s take a closer look at algorithms specifically designed for vehicle analysis. To distinguish between vehicle types, we must first identify where the vehicle is located within an image. Let’s start by examining the algorithm that locates the vehicle. Just as we simplify complex three-dimensional objects into two dimensions for analysis when studying mathematics, converting complex problems into a specific form makes them easier to analyze. Similarly, to detect vehicles in black box footage, we can analyze the video frame by frame, treating each frame as an individual image. This is because we must first locate the vehicle before we can identify its type or specific model. So, how exactly can we detect a vehicle in a photo? Humans, if they know what a vehicle looks like and its characteristics, can relatively easily tell whether a vehicle is present in any given photo. However, computers cannot naturally pick out only the desired objects the way humans do. Therefore, we must train the computer to recognize vehicles by teaching it their shapes and features. In the past, vehicles could be detected using computer vision tools and object detection technologies such as Intel’s OpenCV. Another method involved training a classifier to distinguish between vehicles and non-vehicle objects using positive and negative samples. At that time, by providing separate sets of images of objects to be recognized and those to be ignored, the computer could be trained to learn features from both types of images. In vehicle analysis programs, the training method involved feeding the system thousands or tens of thousands of photos of various vehicle types into the “target” category, while the “non-target” category included a variety of non-vehicle images such as people, dogs, landscapes, backgrounds, the ocean, and mountains. This approach successfully solved the problem of distinguishing vehicle shapes and features. However, determining the exact location of a vehicle within an input image was a different challenge altogether. It was akin to knowing who Mr. Kim was but not knowing which of the eight provinces of Joseon he was in. In the past, a method was used where the pre-trained vehicle shape was divided into a grid of various sizes, and the system checked for the presence of a vehicle by systematically comparing each position and size—from the top-left to the bottom-right of the image—one by one. Although this may seem like a brute-force method, computers can perform repetitive calculations quickly, and various algorithms were developed to reduce the computational load, making this approach practical. However, current vehicle detection technology has advanced significantly beyond these traditional methods. Deep learning-based object detection models analyze the entire image to locate vehicles using bounding boxes and simultaneously determine whether the object is a vehicle or something else. Intel is also currently proposing a method, based on OpenVINO object detection technology, that utilizes deep learning models such as YOLO to detect vehicles and pedestrians in traffic footage.
Now that we have located the vehicle, it is time to examine the algorithms used to identify the vehicle model. In the past, accurately identifying a specific vehicle model was a much more challenging problem than simply locating a vehicle within an image; dimension reduction and feature extraction techniques, such as PCA, were often used to compare features in vehicle images. PCA can help analyze large amounts of vehicle data by highlighting distinct features and representing complex data as a relatively small number of principal components. However, because there is a vast variety of vehicle manufacturers and models—and many vehicles share similar shapes—accurately distinguishing among thousands of different vehicle types requires a significant amount of data and computational power. In particular, even for the same vehicle, images can vary greatly depending on the shooting angle, lighting, weather, the extent to which parts of the vehicle are obscured, and image resolution; therefore, it is difficult to accurately identify the specific model by simply comparing the overall shape. Just as past “celebrity look-alike” apps sometimes identified people who bore no resemblance to a specific individual as look-alikes, methods that compare only certain features of an image have shown limitations in making detailed distinctions. However, this situation has now changed significantly. With advancements in deep learning-based image classification and recognition technologies—which not only locate vehicles but also classify them in detail by manufacturer, model, and body style—systems capable of analyzing real-world vehicles are now being developed. Intel’s OpenVINO resources also provide deep learning-based computer vision technologies for vehicle detection and recognition, and object detection models that identify vehicles and pedestrians in real-time video are now in use. In the past, it required significant computational power to enter all vehicles into a database and compare their distinct features; however, today this process can be automated using deep learning models trained on large-scale vehicle image datasets. Of course, the task of distinguishing vehicle models in detail has not become entirely straightforward. Distinguishing between vehicles with similar appearances, identifying specific vehicle models in low-resolution dashcam footage, or making accurate determinations when a vehicle is partially obscured remain challenging tasks. Nevertheless, vehicle detection and recognition technology has advanced to a degree incomparable to the past and has reached a level where it can be applied to various tasks requiring the automatic analysis of vehicle types captured by dashcams or CCTV. In the future, as more vehicle data is combined with high-performance AI models, technology capable of comprehensively analyzing not only a vehicle’s manufacturer and model but also various features such as color, shape, and license plate is likely to advance even further. Once algorithms with a significant level of vehicle classification capability are reliably developed and deployed in real-world environments, they can help reduce the time and cost required for various tasks that involve identifying vehicle types in dashcam or CCTV footage. Furthermore, vehicle video analysis technology is likely to be utilized more widely in related fields such as accident analysis, traffic management, and criminal investigations.
It will be fascinating to see how vehicle analysis algorithms become even more sophisticated through groundbreaking ideas and technological advancements in the future.
