Many asset tracking applications demand long-lived, low-cost, and continuous monitoring of a large number of items, which has posed a significant challenge to today's RFID design. In order to satisfy these requirements, we propose to adopt transmit-only tags without a receiver, which can offer both low power and low cost. In spite of their great potential, such a platform faces many challenges since it cannot sense the channel, causing the collisions among tag transmissions to be high. It is thus crucial to employ effective multi-user detection schemes at the tag reader to extract valid information from collided signals. Traditional detection schemes, such as successive cancelation, cannot be directly applied to the targeted system. Firstly, due to the simplicity of receiver-less transmit-only tags, there is no mechanism for feedback to the tags that is traditionally needed for accurate multi-user detection. More importantly, these schemes impose serious processing and memory requirements on the underlying system, which makes real-time tracking impossible. In this study, we address these challenges by performing a statistical estimation of the signal amplitude, and by dividing the received signal sequence (from all the tags) and assigning each block to one reader. We also adopt an online learning mechanism so that readers can anticipate the tags that belong to them. We show that the proposed detection algorithm can achieve low detection error under realistic system conditions.