An Optimal Vehicle Counting Framework for Non-CCTV Placements

Automated video surveillance requires CCTV to be strategically placed. However, most available CCTVs are not placed in the most optimal setting for object detection and counting. Most detection and counting algorithm focus on a canonical and vertical view of the road which is easily installed in highways and enables easy counting. However, these methods perform poorly in non-canonical or horizontal high angled CCTVs. Although they provide better coverage, objects are smaller in size with different scales in both traffic lanes. Therefore, this study proposes an optimal counting framework that incorporates guided tiling to help detect smaller objects in horizontal views. An object detection/tracking model is also selected and trained. At the same time, we conduct an ablation study to study the effects of tiling and different counting techniques. Experimental results show that our counting framework can achieve up to 23.2% counting improvement for non-canonical scenes.