ROBUST ALGORITHMS FOR ONLINE VISUAL TRACKING SYSTEM

Abstract

Visual tracking is one of the extremely important research areas in Electro-Optical Tracking System (EOTS) due to its wide-ranging applications. In EOTS, camera is mounted on the set of gimbals to give controlled motion in accordance with target motion. The most relevant application of visual tracking is in the field of defence, where precision tracking of target in hostile conditions with minimal reaction time is of prime importance. For reducing the reaction time, quick gimbal movement and auto detection of potential targets are very much required. Also, fine closed loop tracking is essential for accurate engagement of the target. Therefore, there is a need to develop robust algorithms for quick gimbal movement, target detection, target localization and fine closed loop tracking of fast moving target. In first part of the thesis, introduction of visual tracking and motivation has been presented. It deals with design considerations of visual tracking system and contemporary work. In later part, the thesis deals with the main objectives of this research work formulated as quick acquisition of target, fast movement of gimbaled sight without substantial overshoot, accurate localization of fast moving target in presence of target scaling and fine tracking of accelerating target. Fast steering and quick positioning are prime requirements of current EOTS to achieve quick target acquisition. This thesis puts forward an approach for reducing the reaction time for target acquisition. Algorithm for auto detection of potential targets under dynamic background has been proposed. Also, the design considerations for visual tracking and control system configuration to achieve fast response with proper transient behavior for cued target position have been presented. It has ultimately led to an integrated quick response visual tracking system. Depending on the gimbal configuration, quick movement scheme has been developed. Chapter 3 deals with quick acquisition of target using 2-axis, 2-gimbal sight. In 2-axis and 2-gimbal configuration, error between cue command and gimbal encoder has been modified in such a way that gimbal will follow shortest path and control mode & command will be switched automatically based on instantaneous position error. For larger error, gimbal will be in stabilization mode and it will be steered with maximum permissible rate. During moderate error, rate of gimbal movement will be tapered proportional to the error. Ultimately, gimbal will be switched to position mode for smaller error. In this way, fast gimbal movement has been achieved without incurring high overshoots. Similarly, for fast steering and quick positioning of large field-of-regard, 2-axis & 4- gimbaled sight, slaving of inner/outer gimbals needs to be considered. For steering the lineof- sight (LOS) in the stabilization mode, outer gimbals are slaved to the gyro stabilized inner gimbals. Typically the inner gimbals have direct drives and outer gimbals have geared drives which results in a mismatch of the acceleration capability of its servo loops. This limits the allowable control bandwidth for the inner gimbals. However in order to achieve high i stabilization accuracy, high bandwidth control loops are essential. These contradictory requirements have been addressed in Chapter 4 by designing a suitable command conditioning module for the inner gimbals. Also, large LOS freedom in pitch axis is required to provide a wide area surveillance capacity for airborne application. This leads to a loss of freedom along the yaw axis as the pitch angle goes beyond 70 or so. This is addressed by making the outer gimbal master after certain pitch angle. Moreover, a mounting scheme for gyro has been proposed to accomplish yaw axis stabilization for 110 pitch angle movement with a single 2-axis gyro. Command modifier has been further improved for starting and settling phase of transient response for larger commands. Here, rate limit of command modifier has been changed dynamically based on absolute position error and direction of motion. Achieving smaller slaving error and smooth transient response has paved the way for smaller size windows on the outer gimbal and is helpful in quick target acquisition. As a spin off, this scheme has been employed for quick target cueing and auto-locking of target. It has reduced the reaction time and improved the automation in the operation of the sight. This scheme has been experimentally verified on two tracking systems with different gimbal configurations. Camera zoom operation and fast approaching/receding target causes scaling of acquired target in video frames. Also, fast moving target manifests in large inter frame motion. In general, non-uniform background degrades performance of tracking algorithms. FFT based Correlation algorithms improves tracking in this scenario, but their applications is limited to small inter frame motion. Increasing search region has implication on execution speed of the algorithms. Rapid target scaling, non-uniform background and large inter frame motion of target hinder accurate and long term visual tracking. In Chapter 5, these challenges have been tackled for extended target tracking by augmenting fast Discriminative Scale Space Tracking (fDSST) algorithm with probable target location prediction and target detection. In this modified fDSST, localization of fast motion has been achieved by applying fused outputs of Kalman filter and quadratic regression based prediction before applying fDSST. It has helped in accurate localization of fast motion without increasing search region. In each frame, target location and size have been estimated using fDSST and further refined by target detection near this location. Target detection near fDSST estimated location has helped in acquiring accurate size and location. Smoothing & limiting of trajectory and size of detected target has enhanced tracking performance. Experimental results show considerable improvement of precision, success rate and centre location error tracking performance against state of the art trackers in stringent conditions. Automatic Video Tracker gives tracking error after processing video frames. Since, frame rate is limited to 25 Hz or 30 Hz, tracking error update rate is also limited to this order. In addition, there is a processing delay of about one frame in calculating target movement. Due to these factors, track loop bandwidth is limited to 2 5 Hz. This limits tracking performance against fast moving and accelerating targets. To meet this requirement new control scheme has been proposed in Chapter 6, where dynamics between target and tracking system has ii been modeled as LOS rate, caused by disturbance. This rate has been estimated using Disturbance Observer and used to augment conventional tracking loop output. Combined signal has been used to steer LOS of tracking system and keep target at aimpoint mark. This has resulted in a highly accurate following of the fast moving target. Robust stability analysis of proposed scheme has also been presented. Due to FOV compensation, this scheme is generalized in nature and fine tracking has been achieved with continuous zoom operation of camera. Also, during target acquisition, rate command for stabilization loop has been dynamically restricted to move the camera with permissible rate and avoid motion smearing. Above proposed algorithms have been developed in a systematic way. After problem analysis, schemes were modeled in MATLAB/Simulink and performances were evaluated. In next step, real-time hardware implementation has been done to check their efficacy in real operational scenarios. Finally, proposed schemes have been evaluated in laboratory and field conditions.