A Light Weight Multi-Head SSD Model For ADAS Applications
Chun-Yu Lai, Bo-Xun Wu, Tsung-Han Lee, Vinay Malligere Shivanna, Jiun-In Guo · 2020
Moving objects detection is considered as one of the prime safety indicators in the Advanced Driver Assistance System (ADAS). For implementing on resource-limited embedded platforms and still yield sufficient frame rate and quality, the paper proposes a lightweight multi-head single shot detector (SSD) model that strengthens the moving object detection significantly. The paper also introduces focal loss method to deal with imbalance problem of detecting pedestrians and bikes in training datasets (vehicles, bikes, and pedestrians). Lastly, the proposed lightweight network can be deployed on low-power embedded devices to achieve real-time processing performance (512x256) yielding 30fps.