Frequency domain background modelling and tracking using fractional order gain kalman filter in traffic videos

Abstract

With the advancement of technology the number of vehicles on road are increasing. Roads are more prone to traffic congestion which leads to accidents. It is important to monitor the traffic congestion at crossings for regulating the driver behaviour and preventing the accidents. Vehicle detection and tracking is a key technology for public safety, intelligent transport systems (ITS) and for efficient management of traffic. It is used for traffic surveillance, highway monitoring, highway planning, autonomous car and so on to name a few applications.There are many approaches adopted for vehicle detection and tracking. Selecting a suitable approach for pedestrian and vehicle detection is a challenging task. While evaluating the object detection algorithms, many factors should be considered in order to cope with unconstrained environments, non stationary background, different object motion patterns and the variation in types of objects being detected.The foremost step in vehicle detection and tracking is background extraction. Background modelling is basically extracting stationary objects from the video. Background modelling is a challenging task in vehicle detection and tracking. Accurate background modelling,computational complexity, memory requirement and accurate vehicle segmentation from image sequences are very important in any background subtraction method. In light of this, one of the aims of this thesis is solving problem of vehicle detection under varying illumination conditions. We propose an efficient method for modelling stationary objects. In the proposed method, the background image is modelled in the frequency domain using curvelet transform. The proposed method detects moving vehicles in two steps, i.e. background modelling and foreground object extraction. The foreground object is extracted using morphological operations. As real-time tracking systems rely on the accuracy of the system, the thesis proposes a method for vehicle tracking. This thesis also describes a real time tracking approach for non-linear systems. The moving vehicle is extracted from the image sequences by subtracting the image from the modelled background. In some situation where the motion model may divert from the actual motion model, fractional Kalman lter may diverge. To handle such situations the thesis proposes a method for vehicle tracking using fractional Kalman lter with better accuracy. The non-linearity of the system is linearised using Jacobian. The impact of behaviour of vehicle on error covariance has been reduced using modi ed transition matrix. The fractional states are calculated using Gr unwald-Letnikov (GL) de nition. In some scenarios where the vehicle accelerate or decelerate suddenly or changes direction, the abrupt variation in object state is observed. This thesis also attempts to track such abrupt variations in object state. Kalman lter is widely used for optimal state estimator in object tracking. With known noise and system parameters, Kalman lter tends to stabilize the gain. During sudden transitions, constant gain Kalman lter may diverge.This work proposes a modi ed steady state gain of Kalman lter by introducing fractional feedback loop across Kalman gain. The modi ed Kalman gain is estimated by minimizing the cost function. This modi cation signi cantly improves the accuracy and robustness of the Kalman lter. In order to test the efficacy of the proposed method various online datasets and in-house data has been tested and compared with the algorithms reported in literature. The results indicate improved performance of the proposed algorithm.