With increased security demands, more and more video surveillance systems are installed in public places, such as schools, stations, and shopping malls. Such large-scale monitoring requires 24/7 video analytics, which cannot be achieved purely by manual operations. Thanks to recent advances in artificial intelligence (AI), deep learning algorithms enable automatic video analytics via smart devices, which interpret people/vehicle behaviours in real time to avoid anomalies effectively. Among various video analytical tasks, people search is one of the most critical use cases due to its wide application scenarios, such as searching for missing people, detecting intruders, and tracking suspects. However, current AI-powered people search is generally built upon facial recognition technique, which is effective yet may be privacy-invaded. To address the problem, person re-identification (ReID), which aims to identify person-of-interest without facial information, has become an effective panacea. Despite considerable achievements in recent years, person ReID still faces some tough challenges, such as 1) the strong reliance on identity labels during feature learning, 2) the tradeoff between searching speed and identification accuracy, and 3) the huge modality discrepancy lying between data from different sources, e.g., RGB image and infrared (IR) image. Therefore, the research interest of this thesis is to focus on the above challenges in person ReID, analyze the advantages and limitations of existing solutions, and propose improved solutions for each challenge. Specifically, to alleviate the identity label reliance during feature learning, an improved unsupervised person ReID framework is proposed in Chapter 3, which refines not only imperfect cluster results but also the optimisation directions of samples. Based on the unsupervised setting, we further focus on the tradeoff between searching speed and identification accuracy. To this end, an improved unsupervised binary feature learning scheme for person ReID is proposed in Chapter 4, which derives binary identity representations that not only are robust to transformations but also have low bit correlations. Apart from person ReID conducted within a single modality where both query and gallery are RGB images, cross-modality retrieval is more challenging yet more common in real-world scenarios. To handle the problem, a two-stream framework, facilitating person ReID with on-the-fly keypoint-aware features, is proposed in Chapter 5. Furthermore, the thesis spots several promising research topics in Chapter 6, which are instructive for future works in person ReID