The availability of low-cost video cameras and digital media storage has invited huge investments in developing state-of-the-art algorithms that automatically evaluate and understand video datasets. One such class of algorithm is object tracking which analyzes the data and automatically determines the location of the object in a video sequence. As these algorithms are a prelude to a higher level decision making algorithms, therefore estimation of the trajectory of the object must be accurate and robust under many challenging constraints. A very popular class of object tracking algorithm is the hybrid object tracking category based on integrating Meanshift (MS) and Particle Filter (PF) (MSPF). The purpose of this integration was to overcome the limitation of the PF methods that required a large number of samples/particles PF method to approximate the object state. Consequently, this integration uses the MS optimization procedure to move fewer particles, in the direction of gradient ascent, which represents the dynamics of the target more accurately. The existing methods employ a pre-determined combination of features, inherently assuming that the background would not change. However in uncontrolled environment, it is difficult to specify the background of the object in advance as it moves around the field of view of the camera and thereof this assumption may not often hold. Moreover, hybrid tracking systems based on the MSPF methodology are very compute intensive and it is desirable to reduce this complexity. In the first part of this research, the dissertation aims to investigate an adaptive multi-feature framework that is implemented on top of the MSPF methodology that tracks the object in the local perspective. Essentially that takes care of the dynamic and changing characteristic of the background, which is one of the most important challenges in the object tracking domain. In this research work, an Adaptive Multi-Feature framework is proposed and implemented on top of the MSPF methodology (AMF-MSPF). An adaptive ranking module is proposed that is triggered after a certain criteria is violated, in which case a new set of features are selected for tracking the object. The top ranked features are selected to represent the object, which gives the tracker the ability to adapt to locate the object with an upgraded set of feature. Consequently, this improved local discrimination of the target from its immediate neighboring pixels. In most applications a small portion of computational resources are dedicated to trackers and rest is reserved for higher level decision making tasks, which mandate trackers to be efficient and less complex. Thereby, the second part of the dissertation looks into the complexity of the MSPF methodology. As the MSPF methodology is already a computationally intensive processing task, implementing a feature ranking module on top of it might complicate matters. The feature ranking module also requires a significant portion of the power, thereby a novel MS technique is proposed to free some resources for the ranking module. This novelty comes from an observation that only a fraction of random samples were required by the MS optimization to approximate the similarity metric without inducing significant error. This computational reduction would be advantageous given the complex integration of the MS and PF, because the MS procedure is directly proportional to the number of particles that would take many MS iterations to converge. The proposed novelty in the MS method has reduced its complexity that has greatly impacting the overall complexity of the proposed AMF-MSPF. The proposed AMF-MSPF framework is tested on sequences from the CAVIAR datasets such as Browse and Walkbyshop1and an s8 sequence was taken from the PET dataset. These datasets are known for a number of challenging constraints such as abrupt intensity variations, full occlusions, cluttered background etc. Qualitative results have shown robust and accurate tracking under stringent constraints. In the quantitative analysis, a comparison with the existing methods has been carried out. The proposed framework has shown significant improvements in terms of root mean square error (RMSE), false alarm rate (FAR), and F-SCORE. The average RMSE, FAR, and F_SCORE, over all the video sets, of the proposed AMF-MSPF are 8.68, 0.15, and 0.92, which has improved manifold as compared to the chosen reference methods. Experimental results have proved the effectiveness of the proposed framework.