Article 9 – Automatic Vehicle Recognition Information Systems Using Speed-Up Robust Features

Samuel Mahatmaputra T.

The proposed work aim to implement non learning computer vision method using features detection in real live traffic condition to recognize vehicle by its physical features. The algorithm used in the system is speed up robust features that is invariant to both scaling and rotation of an image. The use of speed up robust features technique is based on its capability to correctly recognize images with different sizes and various angles of rotation. The system was tested against several conditions including tested towards finding moving model and stationed object from the observed video. The accuracy result is promising with an average of 90% correctness with an average performance of 750 millisecond per frame. The work as well aimed to provide conclusion whether the use of scale and rotation invariant technique such as speed-up robust features could contribute positively in the area of vehicle recognition systems

The increasing needs of intelligent traffic monitoring systems [1] as a result of limitation in the manual process of monitoring the video surveillance systems are imminence. Several approaches were proposed to extend computational functionality from historical data and information provider to advance intelligence aggregation and prediction. Many of which use learning method from a large scale of samples to a non-learning method that uses special features for identification. The needs to identify vehicles in the traffic management system and law enforcement process was presented as extremely important [1]. The process involves identifying the type of vehicles by comparing them with vehicles’ pictures database. The proposed work aim to implement non learning computer vision method using features detection in real live traffic condition to recognize vehicle by its physical features.

The algorithm used in the proposed system is speed up robust features [2] that is invariant to both scaling and rotation of an image. The use of speed up robust features technique is based on its capability to correctly recognize images with different sizes and various angles of rotation. The output of the proposed work is an application that could recognize certain vehicle and could match the vehicles within a set of images or video frames. The decision as well was based on its superior performance over the original Lowe’s scale invariant features [3]. The application took an image of a vehicles as input, perform comparison with a set of images and concluded whether particular vehicle exist in the searching domain. The system was tested in real live traffic video surveillance in order to conclude its capability and limitation in the presence of exception conditions in live traffic context

The implementation of speed-up robust features in vehicle recognition system used in this work is divided into several steps describes as follows:

Capturing key bitmap and transforming it into 2D matrix.
Implementing frame grabber to capture each frame of the live traffic video.
Applying speed up robust features to both key bitmap and the observed video frame.
Comparing the descriptors of key bitmap to the descriptors of the observed frame.
Draw bounding box to the location where the maximum descriptor’s similarities occurred.

The experiment’s results were implemented towards 168×177 pixels image resolution that performed detection system within range of 687 to 1183 millisecond per frame. These performances still have room for improvement to keep it below five hundred millisecond per frame. It could be done by implementing a more efficient algorithm towards the frame grabber function, increase the hardware specification of the system, or by lowering the threshold of acceptable number of neighbouring features in the algorithm’s descriptors. However by decreasing the number of features considered in the descriptors will compromise the accuracy level of the system. The hardware that supports the proposed system was: Processor 2.4 GHz Intel Core i5, memory 4 GB 1333MHz DDR3 and Windows operating systems.

The accuracy of both cases is described in table-1 ranging between 90% to 99%, which is acceptable. These satisfactory accuracy results were implemented towards avi video format of the observed road video frames with resolution of 640×480 pixels. The considerably low-resolution video was purposely chosen for the experiment to lower down the searching space during features extraction process.

Please refer to the following paper for further detail:

https://ieeexplore.ieee.org/abstract/document/8843812