In Internet of Vehicles, latency-sensitive object recognition tasks—such as autonomous driving, pedestrian detection, and traffic congestion identification—face challenges when relying on cloud-based computation due to the need for rapid recognition. To address this, we propose a Data Partitioning Method (DDR) based on dynamic programming strategy and receptive field calculation. We have developed a working device selection algorithm based on service utility and device computing power, partitioning the entire network into multiple fusion blocks for collaborative computation and inference optimization within each block. Experimental results demonstrate that our proposed algorithm can reduce accuracy loss to within 5% under strict task latency constraints and largely improve inference speed compared to the state-of-the-art methods.