Accurate and efficient vehicle detection framework based on SSD algorithm

Abstract

Vehicle detection plays an important role in intelligent transportation systems and security. Using the original Single Shot MultiBox Detector (SSD) directly for vehicle detection, lacks accuracy and stability. Moreover, most of the state-of-the-art methods need cost a lot of time to inference. Vehicle detection is often used in complex traffic environments. Therefore, faster detection speed and higher detection accuracy are required. This study is aimed at developing a trade-off between accuracy and speed vehicle detection framework based on the SSD algorithm. To improve the multi-scale detection performance of SSD, semantic information, detailed features and receptive fields are combined to propose the feature pyramid enhancement strategy (FPES). On the other hand, the cascade detection mechanism is proposed to strengthen the positioning capability of SSD and an adaptive threshold acquisition method for object detection module (ODM) stage to improve model accuracy. Finally, a more efficient convolutional network is deployed through network slimming. Experimental results demonstrate that the proposed framework achieves state-of-the-art performance on UA-DETRAC and Udacity benchmarks. Interestingly, the inference time is the lowest for the proposed method than the state-of-the-art methods, promising its application for fast and effective vehicle detection.