Accelerated detection and classification for surveillance applications
Object detection and classification for surveillance applications via deep neural networks have attracted a lot of interests in computer vision (CV) communities. Accurate and fast CV algorithms can alleviate intensive manual labour and reduce human errors due to fatigue and distraction. In detection problem, the aim is to determine bounding boxes which contain interested objects and classify the category of the detected object. Thus, the detection problem can be formulated as a regression problem to localize multiple objects within a frame. Due to very limited computational budgets on the edge devices, server-side solutions like YOLO and R-CNN are not suitable for embedded devices or high-throughput applications that scale to thousands of cameras. It is challenging to achieve real-time object detection performance while maintaining high accuracy. In this research proposal, we focus on reducing computation latency by developing models with a smaller number of trainable parameters to accelerate object detection and classification. First, we take the advantage of the depth separable convolution layer which has less model complexity. Second, we consider hierarchical processing units to localize multiple objects in one-time forward pass of the neural network.