EfficientLiteDet: a real-time pedestrian and vehicle detection algorithm

Since safety plays a crucial role and the top priority, in both unmanned and driver-assistance driving systems, there is a need of efficient and accurate detection of captured objects by object detection algorithms in real-time. Directly applying existing models to tackle real-time pedestrian and vehicle detection tasks captured by high speed moving vehicle scenarios has two problems. First, the target scale varies drastically because the vehicle speed changes greatly. Second, captured images contain both tiny targets and high density targets, which brings in occlusion between targets. To solve the two issues, an efficient light weight real-time detection algorithm is proposed, which is referred to as EfficientLiteDet. Based on Tiny-YOLOv4, one more prediction head is introduced in the proposed model to detect multi-scale targets effectively. In order to detect tiny and occluded denser targets, we used Transformer Prediction Heads (TPH) instead of original anchor detection heads in our model. To explore the potential of self-attention mechanism in TPH, the proposed model integrates “convolutional block attention model” to locate crucial attention region on scenarios with denser targets. Further to improve the detection performance of our model, we applied various data augmentation strategies such as mosaic, mix-up, multi-scale, and random-horizontal-flip during the model training. Extensive experiments are conducted on five challenging pedestrian and vehicle datasets shows that the EfficientLiteDet model has better performance in real-time scenarios. On Pascal Voc-2007, Highway and Udacity datasets, the proposed model achieves mean average precision (mAP) 87.3%, 80.1% and 77.8%, respectively, which is quite better than Tiny-YOLOv4 state-of-the-art algorithm by + 2.4%, 1.8% and + 2.4%, respectively.