Vehicle Speed Detection and Safety Distance Estimation Using Aerial Images of Brazilian Highways

Abstract

In this study, the computational development conducted was based on Convolutional Neural Networks (CNNs), and the You Only Look Once (YOLO) algorithm to detect vehicles from aerial images and calculate the safe distance between them. We analyzed a dataset composed of 896 images, recorded in videos by a DJI Spark Drone. The training set used 60% of the images, 20% for validation, and 20% for the tests. Tests were performed to detect vehicles in different configurations, and the best result was achieved using the YOLO Full-608, with a mean Average Precision(mAP) of 95.6%. The accuracy of the results encourages the development of systems capable of estimating the safe distance between vehicles in motion, allowing mainly to minimize the risk of accidents.

Vision‐based vehicle speed estimation: A survey

“…Although the first approach based on learning was found in 2007 [19], using Boosting type methodologies (Adaboost, also used in [120], which is the only one not based on deep learning), it is not until 2017 that we see consolidated the use of this type of methodologies in all the application domains, including Faster R-CNN [85,88,90,98,101,106,128] an the extended version Mask R-CNN [89], SSD (Single Shot MultiBox Detector) [100,128] and different versions of the YOLO (You Only Look Once) detector, i.e., YOLO [115], YOLOv2 [128,129] and YOLOv3 [104,105,126]. In [91] transfer learning is proposed to fine-tune a vehicle detection model for their specific traffic scenario.…”

“…This way the images can be transformed into a bird's eye view (BEV), which is also known as image warping, in which pixel displacements can be directly transformed to real-world distances. This approach is the most commonly used for both fixed systems [15,16,18,20,22,34,37,52,53,61,66,94,98,99,110] and drone-based [25,30,57,62,64,69,96,[104][105][106]126]. As depicted in Fig.…”

“…The camera settings have a direct impact on the accuracy of the speed estimation method. The first categorization considered is the camera position, including drone cameras [25, 30, 57, 62, 64, 69, 96, 104–106, 126], and traffic cameras that can be classified roughly according to their height as distant () or close () traffic cameras. In some cases, the cameras are located on the side of the road, providing side views of the vehicles [13, 21, 26, 33, 39, 73, 112, 118].…”

“…This way the images can be transformed into a bird's eye view (BEV), which is also known as image warping, in which pixel displacements can be directly transformed to real‐world distances. This approach is the most commonly used for both fixed systems [15, 16, 18, 20, 22, 34, 37, 52, 53, 61, 66, 94, 98, 99, 110] and drone‐based [25, 30, 57, 62, 64, 69, 96, 104–106, 126]. As depicted in Figure 10, and explained in Section 4.2, the computation of the homography involves the calibration of the intrinsic and extrinsic parameters of the camera and previous knowledge of the real dimensions of some part of the scene.…”

“…The first categorization is directly related to the attempt to devise an accurate system to be used for speed enforcement ( speed camera application domain) [13, 17, 18, 32, 33, 39, 41–43, 51, 53, 54, 61, 65–68, 70, 72, 74, 76–78, 80, 82, 84–86, 94, 100–102, 107, 108, 111, 112, 114, 119, 122–125] or a less accurate approach that can still be used for traffic speed estimation ( traffic camera application scenario) [12, 14–16, 19–31, 34–38, 40, 44–50, 52, 55–60, 62–64, 69, 71, 73, 75, 79, 81, 83, 87–93, 95–99, 103–106, 109–111, 113, 115–118, 120, 121, 126–129]. Although in some cases it is claimed that accurate speed detection is targeted, some aspects such as the low resolution of the camera, a limited experimental evaluation or the absence of an appropriate ground truth make us to categorize these manuscripts as traffic speed systems.…”