Automated Vehicle Recognition with Deep Convolutional Neural Networks

Abstract

In recent years there has been growing interest in the use of nonintrusive systems such as radar and infrared systems for vehicle recognition. State-of-the-art nonintrusive systems can report up to eight classes of vehicle types. Video-based systems, which arguably are the most popular nonintrusive detection systems, can report only very coarse classification levels (up to four classes), even with the best-performing vision systems. The present study developed a vision system that can report finer vehicle classifications according to FHWA’s scheme and is also comparable to other nonintrusive recognition systems. The proposed system decoupled object recognition into two main tasks: localization and classification. It began with localization by generating class-independent region proposals for each video frame, then it used deep convolutional neural networks to extract feature descriptors for each proposed region, and, finally, the system scored and classified the proposed regions by using a linear support vector machines template on the feature descriptors. The precision of the system varied by vehicle class. Passenger cars and SUVs were detected at a precision rate of 95%. The precision rates for single-unit, single-trailer, and double-trailer trucks ranged between 92% and 94%. According to receiver operating characteristic curves, the best system performance can be achieved under free flow, daytime or nighttime, and with good video resolution.

Deep learning methods in transportation domain: a review

“…The model achieved 96.05% accuracy on test data. Adu‐Gyamfi et al [44] proposed a deep convolutional neural network in their vision system to detect and classify vehicles into seven classes, and achieved average recall rates between 89 and 99% for seven classes of vehicles.…”

Development of the modified methods to train a neural network to solve the task on recognition of road users

“…Article [19] describes a method of «an end-to-end» recognition of moving objects in a video stream in real time, their classification according to the specified categories in line with the properties based on images and their further monitoring. Recognition of moving objects is carried out by using a pixel difference between consecutive frames of an image.…”

EspiNet V2: a region based deep learning model for detecting motorcycles in urban scenarios

“…Once the object is detected, for instance using GMM, the features extracted using the CNN model (e.g., AlexNet), are used to perform classification [33]. Instead of background subtraction, object localisation uses selective search as in [34]. Nevertheless, the work in [35] proposes a straightforward CNN for detecting and classifying motorcycles.…”