To Turn or Not To Turn, SafeCross is the Answer

Blind area has plagued drivers’ safety ever since the dawn of automobiles. Thanks to the fast-growing vision-based perception technologies, autonomous driving systems can monitor the driving circumstance through a 360-degree view, and hence most blind areas can be avoided. However, in the left turn scenario at an intersection, the opposite road may be blocked by another vehicle parking at the same intersection (see Fig. 1), and in this case, the blind area cannot be observed by the onboard perception module of the autonomous vehicle. A potential fatal collision may occur if the autonomous vehicle turns left while a vehicle is running through the blind area. In this paper, we propose Safecross, a framework that oversees an intersection and delivers blind area warnings to the left-turn vehicles at the intersection if running vehicles are detected in the blind area. In order to provide accurate and reliable real-time warnings in all possible weather conditions, the architecture of Safecross has four major components: video pre-processing (VP) module, video classification (VC) module, few-shot learning (FL) module, and model switching (MS) module. Especially, the VP and VC modules will train a basic model to identify the blind area when a blocking vehicle appears at the intersection. Since the range of the blind area varies in different weather conditions, the FL and MS modules can adapt the basic model to the new condition in real-time to make the blind area identification more accurate. Intuitively, if the blind area is identified timely and accurately, the left-turn throughput of the intersection can be maximized. We have conducted extensive experiments to evaluate our proposed framework. The experiments are performed on a total of 2855 video segments with a time span of 180 days, including sunny, rainy, and snowy weather conditions. Experimental results show how Safecross can guarantee the vehicle’s safety while increasing the left-turn traffic throughput by 50%.