Geometric Deep Particle Filter for Motorcycle Tracking: Development of Intelligent Traffic System in Jakarta

Abstract

-Intelligent Transportation Systems (ITS) is the combination of transportation systems with Informationand Communication Technology (ICT). In Jakarta traffic, there is unique issue that does not arise in developed countries: very large number of motorcycles. Nevertheless, the enabling technologies for the detection, measurement, recording, and information distribution of motorcycle have not been fully developed in the existing researches. With the above considerations, we establish research which aimed to develop enabling technology especially in here for tracking motorcycle using camera. This paper is presented our proposed tracker which called as Geometric Deep Particle Filter (GDPF) for tracking motorcycle using camera. The tracker is inspired by human visual perception which has nonretinotopic nature.Based on particle filter approach, our goal is to improve the transition model in order to overcome motorcycle maneuver. We will exploit this curved nature of the state space using geometric computing theory, such as Lie groups, and Lie algebras. A number of experiments have been conducted for this research, and it has been found that GDPF has achieved certain degree of success in object tracking.

Dynamic swarm particle for fast motion vehicle tracking

“…Several previous studies have used various observation models such as the raw grayscale, raw color, haar-like feature, HOG, and HOG plus raw color. Therefore, several previous studies have utilized deep learning in developing object tracking algorithms, such as Wang and Yeung [17], Gunawan and Jatmiko [18], Zhang et al [19], and Ma et al, [20]. A survey conducted by Wang et al [7] showed that more complex feature extraction methods produce more accurate tracking.…”

“…According to [23], the affine representation can track the object's shape more precisely compared to the representation by vector position. For our proposed algorithm, DSP, we utilize affine representation based on Lie group Aff (2), which is similar to our previous research [18], where this approach was called the geometric transformation method. Therefore, we use 2D affine groups as the target object representation to anticipate these transformations.…”

“…Note that the detail of the derivation of the geometric approach can be seen in [18]. In addition, it is easier to control their parameters.…”

“…This formula has already been implemented in our previous research [18] as the transition model of the visual object tracking algorithm.…”

“…An important part of the observation model is the feature extractor, which extracts the characteristics of a raw image into a more informative representation. In this research, we adopt our version of the approach from our previous research [14,18] by adding the extreme learning machine layer to enhance the speed of the observation model ( Figure 10). The confidence of each particle will determine the particle's best position Xp n i and global best position X n g , as shown in (5).…”