The performance of wireless network diversity multiple access protocols (NDMA) critically depends upon the correct detection of the active users (or the correct estimation of the collision multiplicity). If one or more transmitting users are not correctly detected as active then all the colliding packets will be probably lost due to either a lack of retransmission diversity or an inaccurate equalization design. On the other hand, if some idle users are incorrectly detected as active then extra non-required retransmissions may degrade access-delay. To mitigate this problem, in this paper, we improve the training-based detection method originally proposed for NDMA through two new user detection schemes. Our first proposed scheme is particularly useful in channels with deep and long-term fades, and in systems with low traffic loads. It consists of a set of cooperative nodes that independently collect information about the collision multiplicity and relay it to the base station (BS). The BS finally combines all the contributions with its own collected information in order to improve detection performance. The cooperative nodes only have to extract and relay the header of their received packet, which is less demanding than other cooperative relaying mechanisms where the complete packet has to be processed and/or relayed. The second scheme consists of a sequential detection approach where the BS exploits the information contained in all the received retransmissions. This is different to the conventional NDMA which only considers the information provided by the first collision for user detection. Compared to the cooperative scheme, the second approach is particularly useful in channels with deep and short-term fades and in systems with high traffic loads. The proposed user detection mechanisms are shown to provide an important improvement to the performance of NDMA protocols. Simulation and analytic results are used to assess the benefits of the proposed schemes.