Matching tasks and workers is one of the core problems in spatial crowdsourcing research, but the impact of path planning of workers on task allocation results is usually ignored in the existing literature. There are problems with traditional task assignment methods including slow computing speed, small application scope, and unremarkable collaboration effect. From the perspective of a spatial crowdsourcing platform, this research is oriented toward the spatial crowdsourcing task assignment problem on the road networks and puts forward a QMIX-A* algorithm based on multi-agent reinforcement learning considering workers' path planning. The proposed approach with the minimum completion time of tasks as the objective can shorten the tasks' average completion time, thereby improving users' satisfaction. The effectiveness and stability of the QMIX-A* are verified by a large number of simulation studies. The results of the research can provide decision support for the task allocation and path optimization strategy selection of spatial crowdsourcing service platforms.