Canny algorithm is one of the most popular edge detection algorithms which are used to extract edge features in images. But because of its complicated calculation, using Canny algorithm to detect edges in real-time on traditional embedded systems becomes a difficult problem. However, current high-end embedded processors which integrate GPGPUs (General Purpose Graphics Processing Units) can provide more computing power for embedded systems. This paper proposed a parallel Canny algorithm on the embedded CPU and GPU heterogeneous and collaborative computing system. Experimental results showed that the proposed parallel Canny algorithm brought nearly 50 times speedup on the heterogeneous embedded systems and it fully met the needs of real-time.