Autonomous vehicles must detect nearby road users quickly and reliably, even when objects are partly hidden. We developed a lightweight camera-based method that adapts its processing to scene difficulty, combines information across video frames, and learns from laser-sensor data during training. It reached 38.7% on the benchmark detection measure and processed 26.2 frames per second without laser sensors during use, showing strong potential for fast, lower-cost vehicle perception.

A trajectory prediction model based on Transfomer has been proposed to address the issue of long-term prediction accuracy in complex traffic environments. Optimizing multi-head attention based on knowledge of the scene context and vehicle position generates interactions between maps and agents, as well as between agents themselves. Its effectiveness has been evaluated on the basis of the outdoor dataset, and higher precision was achieved.